DiffusionPipeline.download() on Windows silently skips all subfolder config.json files with huggingface_hub>=1.22.0

Describe the bug

On Windows, DiffusionPipeline.download() builds the allow patterns for per-component configs with os.path.join:
https://github.com/huggingface/diffusers/blob/b8905b9b0f01/src/diffusers/pipelines/pipeline_utils.py#L1740-L1741
which produces vae\config.json, while hub repo paths always use forward slashes (vae/config.json).

Until recently this was harmless: filter_repo_objects in huggingface_hub matched patterns with fnmatch.fnmatch, which on Windows runs os.path.normcase on both pattern and path and unifies the separators, so the backslash pattern still matched. huggingface_hub 1.22.0 switched it to fnmatch.fnmatchcase to make matching case-sensitive on all platforms (huggingface/huggingface_hub#4435), and fnmatchcase does no normalization. Since that release the backslash patterns match nothing:

>>> import os
>>> from fnmatch import fnmatchcase
>>> os.path.join("vae", "config.json")
'vae\\config.json'
>>> fnmatchcase("vae/config.json", os.path.join("vae", "config.json"))
False  # True on Linux/macOS, and True on Windows with fnmatch.fnmatch (hub <= 1.21.0)

Reproduction

Run on Windows with huggingface_hub>=1.22.0:

import os
from diffusers import DiffusionPipeline

folder = DiffusionPipeline.download("hf-internal-testing/tiny-stable-diffusion-torch", cache_dir="./fresh-cache")
for root, _, files in os.walk(folder):
    for f in files:
        print(os.path.relpath(os.path.join(root, f), folder).replace(os.sep, "/"))

Output: weights are fetched, but text_encoder/config.json, unet/config.json and vae/config.json are missing:

model_index.json
feature_extractor/preprocessor_config.json
scheduler/scheduler_config.json
text_encoder/pytorch_model.bin
tokenizer/merges.txt
tokenizer/special_tokens_map.json
tokenizer/tokenizer_config.json
tokenizer/vocab.json
unet/diffusion_pytorch_model.bin
vae/diffusion_pytorch_model.bin

DiffusionPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-torch", cache_dir="./fresh-cache") then fails, and keeps failing on every retry because the incomplete snapshot is considered fully cached. The same snippet on Linux, or on Windows with huggingface_hub<=1.21.0, also fetches the three config.json files.

Logs

`from_pretrained` failure with the reproduction above (transformers 5.x text encoder hits the missing config first and instantiates a default config, so the error is a shape mismatch):


  File "issue_repro.py", line 3, in <module>
    pipe = DiffusionPipeline.from_pretrained("hf-internal-testing/tiny-stable-diffusion-torch", cache_dir="./fresh-cache")
  File "...\site-packages\diffusers\pipelines\pipeline_utils.py", line 1057, in from_pretrained
    loaded_sub_model = load_sub_model(
  File "...\site-packages\diffusers\pipelines\pipeline_loading_utils.py", line 910, in load_sub_model
    loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
  File "...\site-packages\transformers\modeling_utils.py", line 4369, in from_pretrained
    loading_info = cls._finalize_model_loading(model, load_config, loading_info)
  File "...\site-packages\transformers\modeling_utils.py", line 4545, in _finalize_model_loading
    log_state_dict_report(
  File "...\site-packages\transformers\utils\loading_report.py", line 278, in log_state_dict_report
    raise RuntimeError(
RuntimeError: You set `ignore_mismatched_sizes` to `False`, thus raising an error. For details look at the above report!


When a diffusers-native component is the first one loaded from the cached folder (the original real-world case, a pipeline loaded with `transformer` and `text_encoder` passed in, so the VAE came from the snapshot):


  File "...\site-packages\diffusers\pipelines\pipeline_utils.py", line 1057, in from_pretrained
    loaded_sub_model = load_sub_model(
  File "...\site-packages\diffusers\pipelines\pipeline_loading_utils.py", line 910, in load_sub_model
    loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
  File "...\site-packages\diffusers\models\modeling_utils.py", line 1127, in from_pretrained
    config, unused_kwargs, commit_hash = cls.load_config(
  File "...\site-packages\diffusers\configuration_utils.py", line 418, in load_config
    raise EnvironmentError(
OSError: Error no file named config.json found in directory ...\models--vladmandic--Krea-2-Base-sdnq-hadamard-uint4\snapshots\c8ad74b656917d746124cb1aa3991c93a227c7dd\vae.

System Info

• 🤗 Diffusers version: 0.39.0.dev0
• Platform: Windows-11-10.0.26200-SP0
• Running on Google Colab?: No
• Python version: 3.13.11
• PyTorch version (GPU?): 2.12.0+cu132 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 1.22.0
• Transformers version: 5.14.0.dev0
• Accelerate version: 1.14.0
• PEFT version: 0.19.1
• Safetensors version: 0.8.0
• xFormers version: not installed
• Accelerator: NVIDIA RTX 2000 Ada Generation Laptop GPU, 8188 MiB
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

Who can help?

@sayakpaul @DN6

Start in src/diffusers/pipelines/pipeline_utils.py around lines 1740-1741, where DiffusionPipeline.download() builds the component config allow patterns. Reproduce the download on Windows with huggingface_hub 1.22.0 using the tiny-stable-diffusion-torch example, then verify that text_encoder/config.json, unet/config.json, and vae/config.json are present and from_pretrained succeeds.