sampling

app.py

@@ -131,8 +131,12 @@ def generate(
         vision_x=vision_x,
         lang_x=input_ids,
         attention_mask=attention_mask,
-        max_new_tokens=100,
-        num_beams=1,
+        max_new_tokens=50,
+        num_beams=3,
+        do_sample=True,
+        top_p=0.9,
+        top_k=0,
+        no_repeat_ngram_size=3,
     )
 
     gen_text = tokenizer.decode(