Attribute error for AutoTokenizer when loading HF repo from older transformers versions :)

[nil0x9]

System Info

transformers==4.57.2

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers import AutoTokenizer
HF_MODEL_PATH='some/repo/with/old/transformers/version'
tokenizer = AutoTokenizer.from_pretrained(HF_MODEL_PATH)

Error Msg:

File ~/miniconda3/lib/python3.13/site-packages/transformers/tokenization_utils_base.py:2419, in PreTrainedTokenizerBase._from_pretrained(cls, resolved_vocab_files, pretrained_model_name_or_path, init_configuration, token, cache_dir, local_files_only, _commit_hash, _is_local, trust_remote_code, *init_inputs, **kwargs)
   2416     transformers_version = _config.get("transformers_version")
   2418     if transformers_version and version.parse(transformers_version) <= version.parse("4.57.2"):
-> 2419         if _is_local and _config.model_type not in [
   2420             "mistral",
   2421             "mistral3",
   2422             "voxstral",
   2423             "ministral",
   2424             "pixtral",
   2425         ]:
   2426             return tokenizer
   2428 # Expose the `fix_mistral_regex` flag on the tokenizer when provided, even if no correction is applied.

AttributeError: 'dict' object has no attribute 'model_type'

Expected behavior

load success.

[ehoogeveen-medweb]

I'm guessing that <= should be a >=.

[nil0x9]

I'm guessing that <= should be a >=.

nah, the root cause is that they load a dict config object and then use attribute instead of key to access model_type.

I think hf team should really enhance their code quality, CI and review process.

#42379

[vasqu]

Fix is in #42389

[vasqu]

Sorry about that. This is only happening on local tokenizers and not loading from a repo so it went undetected - let me know if the PR doesn't fix the issue

[chtruong814]

Similarly, if we try to load a tokenizer with local_files_only, it fails. For example

AutoTokenizer.from_pretrained('meta-llama/Meta-Llama-3-8B', local_files_only=True)

This fails with

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/opt/venv/lib/python3.12/site-packages/transformers/models/auto/tokenization_auto.py", line 1149, in from_pretrained
    return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/venv/lib/python3.12/site-packages/transformers/tokenization_utils_base.py", line 2108, in from_pretrained
    remote_files = os.listdir(pretrained_model_name_or_path)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: 'meta-llama/Meta-Llama-3-8B'

This worked fine with previous versions where it would pull from the cache.

[ArthurZucker]

yep that's very weird

[chtruong814]

I created a separate issue for the local_files_only case because it was different enough from this.
#42393

[ArthurZucker]

Ty we'll patch

[Linux-Server]

It's not fixed yet?? @ArthurZucker

[lukehinds]

works for me!