Update app.py

app.py

@@ -14,10 +14,10 @@ from rich.markdown import Markdown
 from rich.traceback import install
 install()  # Enable rich tracebacks for easier debugging
 
-# --- Constants ---
+from huggingface_hub import RepositoryInfo
 
-API_URL = "https://api-inference.huggingface.co/models/"
-MODEL_NAME = "meta-llama/Meta-Llama-3.1-8B"  # Replace with your desired model
+model_info = RepositoryInfo.fetch('huggingface/models/{}/{}'.format(" Meta-Llama-3.1-8B ", MODEL_NAME))
+client = InferenceClient(model_info, token=os.environ.get("HF_TOKEN"))
 
 # Chat Interface Parameters
 DEFAULT_TEMPERATURE = 0.9