Update app.py

app.py

@@ -15,7 +15,7 @@ if not os.path.exists(MODEL_FILE):
     print("Download done!")
 
 print("Loading model...")
-llm = Llama(model_path=MODEL_FILE, n_ctx=4096, n_threads=4)
+llm = Llama(model_path=MODEL_FILE, n_ctx=2048, n_threads=4)
 print("Model ready!")
 
 class ChatRequest(BaseModel):
@@ -32,4 +32,4 @@ async def chat(req: ChatRequest):
 
 @app.get("/")
 async def root():
-    return {"status": "ok", "message": "AI Coder is running!"}
+    return {"status": "ok", "message": "AI Coder is running!"}