Update app.py

app.py

@@ -2,6 +2,7 @@ from peft import PeftModel
 from transformers import LlamaTokenizer, LlamaForCausalLM, GenerationConfig
 import gradio as gr
 from torch.cuda import is_available
+from torch import float16
 
 if is_available():
     options = dict(
@@ -10,7 +11,7 @@ if is_available():
     )
 else:
     options = {
-        "torch_dtype": "float16"
+        "torch_dtype": float16
         #"low_cpu_mem_usage": True,
         #"offload_state_dict": True,
         #"offload_folder": "offload",