transformers - 💡(How to fix) Fix PretrainedConfig.from_pretrained: silent default on missing config.json

System Info

• transformers 5.6+ (verified on 5.8.0; 5.5.4 raised)
• introduced by #36033 ("Large/full refactor of from_pretrained", 2025-03-12)

Description

When PretrainedConfig.from_pretrained(path) is given a local directory with no config.json, it silently returns a default-populated config instead of raising OSError. The return type is the config class itself (not Optional) and the docstring does not document this fallback, so a wrong-path call has no way to surface until much later (model build, weight load, OOM, CI timeout).

Reproduction

import tempfile
from transformers import LlamaConfig

with tempfile.TemporaryDirectory() as d:
    cfg = LlamaConfig.from_pretrained(d)
    print(type(cfg).__name__, cfg.model_type)
    # 5.5.4: raises OSError ("...no file named config.json...")
    # 5.6+:  silently returns LlamaConfig with defaults

(AutoConfig raises ValueError("Unrecognized model") because the factory needs model_type; the silent fallback is on concrete PretrainedConfig subclasses.)

Where it changed

cached_files (hub.py) has long special-cased missing config.json to return None (unchanged since 5.5.4; needed so a Hub model id with a cache miss can fall through to download).

#36033 changed the caller chain in configuration_utils.py:

_get_config_dict   -> (None, kwargs)
get_config_dict    -> ({}, kwargs)
from_dict({})      -> cls()   # populates from defaults; no raise

In 5.5.4 the caller raised at this junction.

Suggested fix

For a local directory input, restore the 5.5.4 caller-side raise. Hub-id fallback semantics can stay.

Note

I confirm this is not a pure code-agent issue.