Update app.py

app.py

@@ -24,7 +24,7 @@ model_name_or_path = "TheBloke/phi-2-GPTQ"
 # For example: revision="gptq-4bit-32g-actorder_True"
 
 config = AutoConfig.from_pretrained(model_name_or_path,trust_remote_code=True)
-config.quantization_config["disable_exllama"] = True
+config.quantization_config["use_exllama"] = False
 
 model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
                                              device_map="cpu",