Update app.py

app.py

@@ -12,10 +12,11 @@ bnb_config = BitsAndBytesConfig(
     )
 
 
-tokenizer = AutoTokenizer.from_pretrained("truongghieu/deci-finetuned", trust_remote_code=True).to(device)
+tokenizer = AutoTokenizer.from_pretrained("truongghieu/deci-finetuned", trust_remote_code=True)
+tokenizer.to(device)
 # Load model in this way if use GPU
-model = AutoModelForCausalLM.from_pretrained("truongghieu/deci-finetuned", trust_remote_code=True, quantization_config=bnb_config).to(device)
-
+model = AutoModelForCausalLM.from_pretrained("truongghieu/deci-finetuned", trust_remote_code=True, quantization_config=bnb_config)
+model.to(device)
 # model = AutoModelForCausalLM.from_pretrained("truongghieu/deci-finetuned", trust_remote_code=True)
 # Move the model to the GPU if available