Update gradio_mix.py

gradio_mix.py

@@ -426,14 +426,15 @@ class MMSAlignModel:
 
 class WhisperxModel:
     def __init__(self, model_name):
-        # Lazily construct the WhisperX pipeline so that on Spaces we only
-        # touch CUDA inside spaces.GPU workers.
+        # Lazily construct the WhisperX pipeline.
         self.model_name = model_name
         self.model = None        
-        if IS_SPACES and torch.cuda.is_available():
-            self.device = "cuda"
-        else:
+        # In HF Spaces ZeroGPU, CUDA/cuDNN may not be fully available for
+        # WhisperX, so we force CPU there. Locally, prefer CUDA if available.
+        if IS_SPACES:
             self.device = "cpu"
+        else:
+            self.device = "cuda" if torch.cuda.is_available() else "cpu"
 
     def _ensure_model(self):
         if self.model is not None:
@@ -673,9 +674,6 @@ def get_transcribe_state(segments):
         "word_bounds": [f"{word['start']} {word['word']} {word['end']}" for word in segments["words"]]
     }
 
-@spaces.GPU
-@torch.no_grad()
-@torch.inference_mode()
 def transcribe(seed, audio_info):
     if transcribe_model is None:
         raise gr.Error("Transcription model not loaded")