fix: only keep stop sequence buffer if we have some

server/text_generation_server/utils/tokens.py

@@ -112,7 +112,7 @@ class StoppingCriteria:
         self.stop_sequence_criterias = stop_sequence_criterias
         self.max_new_tokens = max_new_tokens
         self.current_tokens = 0
-        self.current_output = "test"
+        self.current_output = ""
         self.ignore_eos_token = ignore_eos_token
 
     def __call__(self, last_token: int, last_output: str) -> Tuple[bool, Optional[str]]:
@@ -123,6 +123,7 @@ class StoppingCriteria:
         if not self.ignore_eos_token and last_token == self.eos_token_id:
             return True, FinishReason.FINISH_REASON_EOS_TOKEN
 
+        if self.stop_sequence_criterias:
             self.current_output += last_output
             # There is no need to keep an output that is too long
             if len(self.current_output) > 300: