Add custom handler for TTS inference

handler.py

@@ -70,11 +70,15 @@ class EndpointHandler:
         if isinstance(out, np.ndarray):
             audio = out
         elif isinstance(out, list):
-            # If output is a list, take the first element
-            audio = np.array(out[0]) if len(out) > 0 else np.array(out)
+            # If output is a list, take the first element and handle it
+            first_item = out[0] if len(out) > 0 else out
+            if hasattr(first_item, 'cpu'):
+                audio = first_item.cpu().numpy()
+            else:
+                audio = np.array(first_item)
         elif hasattr(out, 'cpu'):
-            # If it's a tensor, convert to numpy
-            audio = out.cpu().numpy()
+            # If it's a tensor (including CUDA tensors), move to CPU and convert to numpy
+            audio = out.detach().cpu().numpy()
         else:
             # Fallback: try to convert to numpy array
             audio = np.array(out)