Update app.py

app.py

@@ -26,7 +26,7 @@ model = f"./Models/{repo_id}/{filename}"
 max_length = 8192
 pipeline = chatglm_cpp.Pipeline(model, max_length=max_length)
 
-# messages = []
+messages = []
 
 def respond(
     message,
@@ -36,9 +36,9 @@ def respond(
     temperature,
     top_p,
 ):
-    # global messages
+    global messages
 
-    # print(messages)
+    print(messages)
     
     generation_kwargs = dict(
         max_length=max_length,
@@ -51,8 +51,8 @@ def respond(
         stream=True,
     )
 
-    # if messages == []:
-    messages = [chatglm_cpp.ChatMessage(role="system", content=system_message)]
+    if messages == []:
+        messages = [chatglm_cpp.ChatMessage(role="system", content=system_message)]
 
     print(messages)
 
@@ -64,15 +64,18 @@ def respond(
 
     messages.append(chatglm_cpp.ChatMessage(role="user", content=message))
 
+    print(messages)
+
     response = ""
     yield response
     chunks = []
+    yield response
 
     for chunk in pipeline.chat(messages, **generation_kwargs):
         response += chunk.content
         chunks.append(chunk)
         yield response
-
+    yield response
     messages.append(chatglm_cpp.ChatMessage(role="assistant", content=response))
 
     print(messages)