chunked/nltk_chunking/main_classes_pipelines/chunk_23.txt

In that case, the whole batch will need to be 400
tokens long, so the whole batch will be [64, 400] instead of [64, 4], leading to the high slowdown.