Update app.py

app.py

@@ -19,7 +19,11 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 
 # Load LoRA
-model = PeftModel.from_pretrained(model, LORA_PATH)
+model = PeftModel.from_pretrained(
+    model,
+    LORA_PATH,
+    device_map={"": "cpu"}
+)
 model.eval()