Update app.py

app.py

@@ -10,10 +10,8 @@ def clear_memory():
 
 model_name = "GIGAParviz/Firooze_test"
 
-
-bnb_config = BitsAndBytesConfig(llm_int8_threshold=6.0)
 tokenizer = AutoTokenizer.from_pretrained(model_name)
-model = AutoModelForCausalLM.from_pretrained(model_name, quantization_config=bnb_config)
+model = AutoModelForCausalLM.from_pretrained(model_name)
 model = model.to("cpu")
 model.gradient_checkpointing_enable()