Generated with:

```python
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_3b_v2")
assert tokenizer.is_fast
tokenizer.save_pretrained("./output/openlm-research/open_llama_3b_v2")
```

Sidenote: it looks like the issue with fast tokenizer mentioned in the README (https://github.com/huggingface/transformers/issues/24233) has been resolved.