Update app.py

app.py

@@ -141,7 +141,7 @@ def initialize_model():
     model = AutoModelForCausalLM.from_pretrained(
         MODEL_ID,
         device_map="cuda",
-        #quantization_config=quantization_config,
+        quantization_config=quantization_config,
         torch_dtype=torch.bfloat16,
         trust_remote_code=True
     ).to("cuda")