Update inference_gradio.py

inference_gradio.py

@@ -46,7 +46,7 @@ class UVR5:
         # Keep paths as strings; actual model is loaded lazily.
         self.model_dir = str(model_dir)
         self.code_dir = str(code_dir)
-        self.model = self.load_model()
+        self.model = None
         self.device = "cpu"
 
     def load_model(self, device: str = "cpu"):
@@ -56,7 +56,10 @@ class UVR5:
 
         if self.code_dir not in sys.path:
             sys.path.append(self.code_dir)
-        
+            
+        if self.model is not None:
+            return self.model
+            
         from multiprocess_cuda_infer import ModelData, Inference
 
         model_path = os.path.join(self.model_dir, "Kim_Vocal_1.onnx")
@@ -82,6 +85,7 @@ class UVR5:
     def denoise(self, audio_info):
         print("denoise UVR5: ", audio_info)
         # # On Spaces, force CPU; locally prefer CUDA if available.
+        self.model = self.load_model()
         input_audio = load_wav(audio_info, sr=44100, channel=2)
         output_audio = self.model.demix_base({0: input_audio.squeeze()}, is_match_mix=False, device="cpu")
         return output_audio.squeeze().T.cpu().numpy(), 44100