Update handler.py

handler.py

@@ -26,7 +26,7 @@ class EndpointHandler():
         while True:
             prompt = self.default_template.format(char_name = char_name, user_name = user_name, user_input = "\n".join(user_input))
             input_ids = self.tokenizer(prompt + f"\n{char_name}:", return_tensors = "pt").to("cuda")
-            if input_ids.input_ids.size(1) >= 2048:
+            if input_ids.input_ids.size(1) > 2000:
                 chats_curled += 1
                 user_input = user_input[chats_curled*2:]
             else: break