unload model after inference

app.py

@@ -57,6 +57,7 @@ def inference_gradio(input, model_choice):  # input is audio waveform in [T, cha
 
 @spaces.GPU(duration=120)
 def inference_model(audio_input, h, model):
+    # load model to device
     model.to(device)
 
     def get_mel(x):
@@ -69,12 +70,17 @@ def inference_model(audio_input, h, model):
 
         y_g_hat = model(spec_gt)
 
-        audio_gen = y_g_hat.squeeze()
+        audio_gen = y_g_hat.squeeze().cpu()
         spec_gen = get_mel(audio_gen.unsqueeze(0))
         audio_gen = audio_gen * MAX_WAV_VALUE
-        audio_gen = audio_gen.cpu().numpy().astype('int16')
+        audio_gen = audio_gen.numpy().astype('int16')
 
-        return audio_gen, spec_gen[0].cpu()
+    # unload to cpu
+    model.to('cpu')
+    # delete gpu tensor
+    del spec_gt, y_g_hat
+    
+    return audio_gen, spec_gen[0]
 
 
 css = """