trying out longer token count to see if we can get more verbose answer

handler.py

@@ -207,7 +207,7 @@ class EndpointHandler():
             img_features=None,
             img_attn_mask=inputs["img_attn_mask"][:, :-1].to("cuda"),
             use_cache=True,
-            max_new_tokens=128,
+            max_new_tokens=512,
         )
         generated_text = self.processor.batch_decode(generated_ids, skip_special_tokens=True)[0]