Update handler.py

handler.py

@@ -22,5 +22,7 @@ class EndpointHandler():
         instruction = data.pop("instruction",data)
         text = data.pop("text", data)
         inputs = [[s, instruction] for s in text]
+        if (self.device):
+            inputs = torch.tensor(inputs).to(self.device)  # Move inputs to the GPU
         embeddings = self.model.encode(inputs)
         return embeddings.tolist()