Update app.py

app.py

@@ -27,9 +27,8 @@ def load_model():
 
         model = AutoModelForCausalLM.from_pretrained(
             MODEL_ID,
-            device_map="auto",
+            device_map="cpu",
             load_in_4bit=True,
-            torch_dtype=torch.float16,
             low_cpu_mem_usage=True
         )