`register_to_config` mislabels positional `__init__` args as `_use_default_values`, so `from_config` round trips silently revert them to defaults

Inizia in src/diffusers/configuration_utils.py intorno a inner_init alle righe 725-726 ed esamina extract_init_dict intorno alle righe 499-501. Esegui la riproduzione e aggiungi la copertura di regressione in tests/others/test_config.py. Il lavoro è completato quando sia gli argomenti posizionali sia quelli del costruttore con nome preservano i loro valori espliciti attraverso i round trip di from_config.

Describe the bug

@register_to_config computes _use_default_values as set(new_kwargs) - set(init_kwargs), but init_kwargs only contains keyword arguments (src/diffusers/configuration_utils.py:725-726). Any constructor argument passed positionally is mislabeled as "used default value", and extract_init_dict then strips it on every from_config round trip (configuration_utils.py:500-501).

• Expected: per the comment at configuration_utils.py:499 ("Skip keys that were not present in the original config, so default __init__ values were used") and the design intent in https://github.com/huggingface/diffusers/pull/3929#issuecomment-1618919655, _use_default_values should only contain parameters the caller did not provide — positional and keyword calls should round-trip identically.
• Actual: the config displays the explicitly-set value, but from_config(obj.config) (the documented scheduler-swap pattern) silently reverts it to the class default. Affects every ConfigMixin subclass.

I'd be happy to open a PR: exclude positionally-bound parameter names from the _use_default_values computation in inner_init (a two-line change), plus a regression test in tests/others/test_config.py — once a maintainer acks, per the AI-assisted contributions guidelines.

Reproduction

from diffusers import DDIMScheduler, EulerDiscreteScheduler

s = DDIMScheduler(500)                        # positional, explicit non-default value
print(s.config.num_train_timesteps)           # 500
print("num_train_timesteps" in s.config["_use_default_values"])                  # True  <-- mislabeled

print(DDIMScheduler.from_config(s.config).config.num_train_timesteps)           # 1000, expected 500
print(EulerDiscreteScheduler.from_config(s.config).config.num_train_timesteps)  # 1000, expected 500

k = DDIMScheduler(num_train_timesteps=500)    # keyword control group
print(DDIMScheduler.from_config(k.config).config.num_train_timesteps)           # 500, correct

Logs

(no traceback — the failure mode is a silently wrong value)

System Info

• 🤗 Diffusers version: 0.40.0.dev0 (main @ 614ae4b)
• Platform: macOS-26.5.2-arm64-arm-64bit-Mach-O
• Python version: 3.13.3
• PyTorch version (GPU?): 2.13.0 (False)
• Huggingface_hub version: 1.27.0
• Transformers version: 5.15.0
• Safetensors version: 0.8.0
• Using GPU in script?: No

Who can help?

No response

Disclosure: this report was prepared with AI assistance; I reproduced the issue locally and reviewed every claim myself.