chunked/nltk_chunking/model_doc_code_llama/chunk_20.txt

The reason is that the model will first be downloaded (using the dtype of the checkpoints online) and then will be casted to the default dtype of torch (becomes torch.float32).