Add application file

app.py

@@ -56,7 +56,8 @@ else:
         BASE_MODEL, 
         device_map={"": device}, 
         low_cpu_mem_usage=True,
-        load_in_8bit=True,
+        load_in_8bit=False,
+        torch_dtype=torch.float16,
     )
     model = PeftModel.from_pretrained(
         model,