[sebastianvlad1]

What does this PR do?

This PR keeps the default precedence introduced in #45094: explicitly registered local classes continue to override checkpoint auto_map code by default.

It adds an explicit prefer_auto_map=True opt-in to AutoConfig.from_pretrained() for cases where callers want the checkpoint's auto_map code to win instead. AutoModel.from_pretrained(..., prefer_auto_map=True) supports the same workflow when the config is auto-loaded, because the flag is forwarded into AutoConfig.from_pretrained().

The PR also preserves the local config-file-path fix from the earlier revision by normalizing file paths to their containing directory before calling get_class_from_dynamic_module().

Fixes #45698

Code Agent Policy

• I confirm that this is not a pure code agent PR.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
  from_pretrained loads wrong custom module after save_pretrained #45698
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Validation

python -m pytest tests/models/auto/test_configuration_auto.py -k "dynamic_config_conflict or prefer_auto_map or saved_dynamic_config_conflict" -s -v
# passed

python -m pytest tests/models/auto/test_modeling_auto.py -k "dynamic_model_conflict or prefer_auto_map or saved_dynamic_model_conflict" -s -v
# passed

python utils/tests_fetcher.py
# ran

python utils/checkers.py ruff_check,ruff_format,init_isort,sort_auto_mappings --fix
# passed

Who can review?

@Cyrilvallez @Rocketknight1 @nurpax

[Qodo-Free-For-OSS]

Hi, AutoConfig.from_pretrained can now select dynamic loading for a local config file input, but it still passes the file path into get_class_from_dynamic_module, which only treats directories as local and will mis-handle the file path as a Hub repo id.

Severity: action required | Category: correctness

How to fix: Normalize file path to dir

Agent prompt to fix - you can give this to your LLM of choice:

Issue description

AutoConfig.from_pretrained can take a file path (e.g. .../config.json). With this PR’s new condition, the code can enter the dynamic-loading branch, but still passes the file path to get_class_from_dynamic_module(). That dynamic loader only recognizes directories as local (via os.path.isdir), so a file path is treated like a remote repo id and dynamic loading fails.

Issue Context

• _local_checkpoint_contains_auto_map_module() already normalizes file paths to dirname, so the code can decide “local checkpoint contains module”.
• The subsequent dynamic load call should use the same normalized directory.

Fix Focus Areas

• src/transformers/models/auto/configuration_auto.py[91-109]
• src/transformers/models/auto/configuration_auto.py[401-421]

Suggested change

When pretrained_model_name_or_path is a file, compute checkpoint_dir = os.path.dirname(os.fspath(pretrained_model_name_or_path)) and pass checkpoint_dir to get_class_from_dynamic_module() (and likely to config_class.from_pretrained() as well, for consistency with other local-dir loads).

---

Found by Qodo code review. FYI, Qodo is free for open-source.

[sebastianvlad1]

@Rocketknight1 would you mind taking a look when you get a chance?

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto

[Rocketknight1]

cc @hmellor since I haven't investigated this too deeply and you were commenting on the original issue, but if you want me to take it instead, just ping me!

[nurpax]

Sorry for taking so long to review!

I read through the code but tbh, it would be easier for code review if code formatting was applied in a separate commit and the real beef of the change kept separate from that. Similarly, I would've dropped the commit that doesn't relate to prefer_auto_map fix, since that shouldn't be needed anymore, right?

Maybe HF has some policy on when is the right time to apply code formatting changes, maybe it's ok to keep them in the same commit.. but it increases the code review load a fair bit.

With these changes, how does the original repro from #45698 change behavior?

[sebastianvlad1]

Thanks for the review @nurpax!

Yes, the intent here was to implement the direction suggested in #45698: expose the precedence explicitly with a per call prefer_auto_map option, while preserving the current default behavior from #45094.

So the behavior is:

• Original from_pretrained loads wrong custom module after save_pretrained #45698 repro unchanged: behavior remains local first by default. The explicitly registered local class still wins, preserving fix: prefer registered config over remote code in AutoConfig.from_pretrained #45094.
• Original from_pretrained loads wrong custom module after save_pretrained #45698 repro with prefer_auto_map=True on the relevant from_pretrained call: checkpoint auto_map wins, so pretrained_b is loaded from its saved dynamic module and gets the expected Magic B behavior.