Update app.py

app.py

@@ -4,10 +4,11 @@ from pydantic import BaseModel
 from llama_cpp import Llama
 
 
-llm = AutoModelForCausalLM.from_pretrained("TowerInstruct-7B-v0.2.Q8_0.gguf",
-max_new_tokens = 2084,
-threads = 3,
-)
+llm = Llama.from_pretrained(
+    	repo_id="TheBloke/TowerInstruct-7B-v0.1-GGUF",
+    	filename="towerinstruct-7b-v0.1.Q5_K_M.gguf",
+        n_ctx = 4096,
+    )
 
 #Pydantic object
 class validation(BaseModel):
@@ -38,10 +39,4 @@ def translate_zh_to_en(llm, text):
 # <|im_start|>assistant
 @app.post("/translate")
 async def stream(item: validation):
-    llm = Llama.from_pretrained(
-    	repo_id="TheBloke/TowerInstruct-7B-v0.1-GGUF",
-    	filename="towerinstruct-7b-v0.1.Q5_K_M.gguf",
-        n_ctx = 4096,
-    )
-
     return translate_zh_to_en(llm, item.prompt)