Update main.py

main.py

@@ -1,6 +1,4 @@
 from flask import Flask, request, abort
-from flask_ngrok import run_with_ngrok # colab
-from llama_cpp import Llama
 import json
 import os
 
@@ -21,7 +19,6 @@ working_status = os.getenv("DEFALUT_TALKING", default = "true").lower() == "true
 
 app = Flask(__name__)
 
-llm = Llama(model_path="Breeze-7B-Instruct-v0.1-Q8_0.gguf", n_gpu_layers=32)
 line_handler = WebhookHandler(YOUR_CHANNEL_SECRET)
 line_bot_api = LineBotApi(YOUR_CHANNEL_ACCESS_TOKEN)
 
@@ -55,9 +52,7 @@ def handle_message(event):
         return
 
     if working_status:
-        QA = f"HUMAN: {event.message.text}? ASSISTANT: "
-        out = llm(QA, max_tokens=512, temperature = 0.5, echo=True)
-        out = out['choices'][0]['text'].split('ASSISTANT:')[1]
+        out = event.message.text
         if not out:
             out = "不懂您的意思!"
         line_bot_api.reply_message(