add flash attention, reorder examples

app.py

@@ -18,8 +18,8 @@ description = "Test out Shisa 7B in either English or Japanese. If you aren't ge
 placeholder = "Type Here / ここに入力してください" 
 examples = [
     ["What are the best slices of pizza in New York City?"],
-    ['How do I program a simple "hello world" in Python?'],
     ["東京でおすすめのラーメン屋ってどこ？"],
+    ['How do I program a simple "hello world" in Python?'],
     ["Pythonでシンプルな「ハローワールド」をプログラムするにはどうすればいいですか？"],
 ]
 
@@ -40,6 +40,7 @@ model = AutoModelForCausalLM.from_pretrained(
         bnb_4bit_use_double_quant=True,
         bnb_4bit_compute_dtype=torch.bfloat16
     ),
+    use_flash_attention_2=True
 )
 streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)