update

app.py

@@ -22,8 +22,8 @@ python moss_web_demo_gradio.py --model_name fnlp/moss-moon-003-sft --gpu 0,1,2,3
 
 
 import gradio as gr
-from models.hf_qwen2 import bot
-# from models.cpp_qwen2 import bot
+# from models.hf_qwen2 import bot
+from models.cpp_qwen2 import bot
 
 
 #
@@ -77,19 +77,19 @@ def parse_text(text):
 def generate_query(chatbot, history):
     if history and history[-1]["role"] == "user":  # 该生成response了
         gr.Warning('You should generate assistant-response.')
-        return None, chatbot, history
-    streamer = bot.generate_query(history)
-    # chatbot.append((query, ""))
-
-    query = ""
-    for new_text in streamer:
-        print(new_text)
-        query += new_text
-        yield query, chatbot, history
+        yield None, chatbot, history
+    else:
+        chatbot.append(None)
+        streamer = bot.generate_query(history, stream=True)
+        query = ""
+        for new_text in streamer:
+            print(new_text)
+            query += new_text
+            chatbot[-1] = (query, None)
+            yield query, chatbot, history
 
-    chatbot.append((query, None))
-    history.append({"role": "user", "content": query})
-    yield query, chatbot, history
+        history.append({"role": "user", "content": query})
+        yield query, chatbot, history
 
 
 def generate_response(query, chatbot, history):
@@ -103,18 +103,24 @@ def generate_response(query, chatbot, history):
     """
     if query and history[-1]["role"] != "user":
         history.append({"role": "user", "content": query})
+    query = history[-1]["content"]
 
     if history[-1]["role"] != "user":
         gr.Warning('You should generate or type user-input first.')
-        return chatbot, history
-
-    response = bot.generate_response(history)
-    query = history[-1]["content"]
-    chatbot[-1] = (query, response)
-    history.append({"role": "assistant", "content": response})
-    print(f"chatbot is {chatbot}")
-    print(f"history is {history}")
-    return chatbot, history
+        yield chatbot, history
+    else:
+        streamer = bot.generate_response(history, stream=True)
+        response = ""
+        for new_text in streamer:
+            print(new_text)
+            response += new_text
+            chatbot[-1] = (query, response)
+            yield chatbot, history
+
+        history.append({"role": "assistant", "content": response})
+        print(f"chatbot is {chatbot}")
+        print(f"history is {history}")
+        yield chatbot, history
 
 
 def generate():