fixed bug

app_modules/llm_inference.py

@@ -80,7 +80,11 @@ class LLMInference(metaclass=abc.ABCMeta):
         if self.llm_loader.streamer is not None and isinstance(
             self.llm_loader.streamer, TextIteratorStreamer
         ):
-            count = 2 if len(inputs.get("chat_history")) > 0 else 1
+            count = (
+                2
+                if "chat_history" in inputs and len(inputs.get("chat_history")) > 0
+                else 1
+            )
 
             while count > 0:
                 try: