Update app.py

app.py

@@ -190,7 +190,7 @@ def fast_decode_step(token, e_seq, h, c, decoder_model, e_mask=None):
     c = tf.convert_to_tensor(c, dtype=tf.float32)
     return decoder_model([token, e_seq, e_mask, h, c], training=False)
 
-def reply(text, max_decode_len=MAX_LEN, rep_penalty=1.3, beam_width=13, length_penalty=0.7, temperature=0.7):
+def reply(text, max_decode_len=MAX_LEN, rep_penalty=1.3, beam_width=7, length_penalty=0.7, temperature=0.7):
     text = text.strip()
     if MoE:
         if not text.startswith("<TASK_"):