Update app.py

app.py

@@ -61,6 +61,7 @@ model_path = "facebook/opt-1.3b"
 offload_folder = "offload"
 tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
 model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", offload_folder=offload_folder)
+model.to('cpu')
 
 model_q = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", offload_folder=offload_folder)
 model_q.to('cpu')