Update app.py

app.py

@@ -33,11 +33,11 @@ repetition_penalty=1.7
 
 tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
 model = transformers.AutoModelForCausalLM.from_pretrained(model_name,
-    device_map="auto",
-    torch_dtype=torch.float16,
+    device_map="auto"
+    torch_dtype=torch.bfloat16,
 #   load_in_4bit=True
 )
-model.eval()
+# model.eval()
 
 class StarlingBot:
     def __init__(self, system_prompt="The following dialogue is a conversation"):