diff --git a/hf-inference-example.py b/hf-inference-example.py
index 5ce2558..faacf47 100644
--- a/hf-inference-example.py
+++ b/hf-inference-example.py
@@ -1,5 +1,9 @@
 import llamahf
 
+# to save memory use bfloat16 on cpu
+# import torch
+# torch.set_default_dtype(torch.bfloat16)
+
 MODEL = 'decapoda-research/llama-7b-hf'
 # MODEL = 'decapoda-research/llama-13b-hf'
 # MODEL = 'decapoda-research/llama-30b-hf'