Update inference_gradio.py

inference_gradio.py

@@ -82,15 +82,8 @@ class UVR5:
         # main process. The heavy UVR5 loading happens lazily inside
         # @spaces.GPU functions; this guard is kept only for the CPU path to
         # avoid any accidental CUDA init.
-        if IS_SPACES and device == "cpu":
-            orig_is_available = torch.cuda.is_available
-            torch.cuda.is_available = lambda: False
-            try:
-                uvr5_model.load_model(model_path, 1)
-            finally:
-                torch.cuda.is_available = orig_is_available
-        else:
-            uvr5_model.load_model(model_path, 1)
+        
+        uvr5_model.load_model(model_path, 1)
 
         self.model = uvr5_model
         self.device = device
@@ -99,10 +92,9 @@ class UVR5:
     def denoise(self, audio_info):
         # Prefer GPU if available; on Spaces this runs inside @spaces.GPU so
         # CUDA can be safely initialized here.
-        device = "cuda" if torch.cuda.is_available() else "cpu"
-        model = self.load_model(device=device)
+        model = self.load_model(device="cpu")
         input_audio = load_wav(audio_info, sr=44100, channel=2)
-        output_audio = model.demix_base({0:input_audio.squeeze()}, is_match_mix=False, device=device)
+        output_audio = model.demix_base({0:input_audio.squeeze()}, is_match_mix=False, device="cpu")
         # transform = torchaudio.transforms.Resample(44100, 16000)
         # output_audio = transform(output_audio)
         return output_audio.squeeze().T.cpu().numpy(), 44100