From 80a69204c1752fb8d754405b1d63cd345fa85e3e Mon Sep 17 00:00:00 2001
From: OlivierDehaene <23298448+OlivierDehaene@users.noreply.github.com>
Date: Thu, 14 Dec 2023 17:01:43 +0100
Subject: [PATCH] fix: slice stopping criteria buffer

---
 server/text_generation_server/utils/tokens.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/server/text_generation_server/utils/tokens.py b/server/text_generation_server/utils/tokens.py
index ff0556df..53722fec 100644
--- a/server/text_generation_server/utils/tokens.py
+++ b/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 = ""
+        self.current_output = "test"
         self.ignore_eos_token = ignore_eos_token
 
     def __call__(self, last_token: int, last_output: str) -> Tuple[bool, Optional[str]]:
@@ -124,6 +124,10 @@ class StoppingCriteria:
             return True, FinishReason.FINISH_REASON_EOS_TOKEN
 
         self.current_output += last_output
+        # There is no need to keep an output that is too long
+        if len(self.current_output) > 300:
+            # Slice to -200 to avoid doing it all the time
+            self.current_output = self.current_output[-200:]
         for stop_sequence_criteria in self.stop_sequence_criterias:
             if stop_sequence_criteria(self.current_output):
                 return True, FinishReason.FINISH_REASON_STOP_SEQUENCE