Disable forcing CUDA

~550 ms for 100 embeddings on T4

handler.py

@@ -14,5 +14,5 @@ class EndpointHandler():
             A :obj:`list` | `dict`: will be serialized and returned
         """
         sentences = data.pop("inputs",data)
-        embeddings = self.model.encode(sentences, device="cuda", batch_size=100)
+        embeddings = self.model.encode(sentences, batch_size=100)
         return embeddings.tolist()