Update app.py

app.py

@@ -97,6 +97,8 @@ def mode_load(path):
 
 
 @spaces.GPU()
+    
+def stream_chat(message, history: list, temperature: float, max_length: int, top_p: float, top_k: int, penalty: float):
     model = AutoModelForCausalLM.from_pretrained(
         MODEL_ID,
         torch_dtype=torch.bfloat16,
@@ -104,7 +106,6 @@ def mode_load(path):
         trust_remote_code=True
     )
     model.eval()
-def stream_chat(message, history: list, temperature: float, max_length: int, top_p: float, top_k: int, penalty: float):
     print(f'message is - {message}')
     print(f'history is - {history}')
     conversation = []