Offline mode rejects local LoRA files and directories without weight_name

[fropych]

Describe the bug

When HF_HUB_OFFLINE=1 is enabled, or local_files_only=True is passed,
loading LoRA weights from an existing local file or directory without an
explicit weight_name fails with:

ValueError: When using the offline mode, you must specify a weight_name.

Offline mode should prevent HTTP requests, but it should not prevent Diffusers
from inspecting an explicitly provided local filesystem path.

The issue is in _best_guess_weight_name():

diffusers/src/diffusers/loaders/lora_base.py

Lines 277 to 290 in 1aadc65

	def _best_guess_weight_name(
	pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
	):
	if local_files_only or HF_HUB_OFFLINE:
	raise ValueError("When using the offline mode, you must specify a `weight_name`.")
	targeted_files = []
	if os.path.isfile(pretrained_model_name_or_path_or_dict):
	return
	elif os.path.isdir(pretrained_model_name_or_path_or_dict):
	targeted_files = [f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)]
	else:
	files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings

The function checks local_files_only / HF_HUB_OFFLINE before checking
os.path.isfile() and os.path.isdir(). As a result, existing local paths are
rejected before they can be inspected.

I understand that #6184 intentionally requires an explicit weight_name for a
remote repository ID in offline mode. Without access to model_info(),
Diffusers cannot automatically determine which file exists in a Hub repository.
This report does not propose changing that behavior.

This is related to #6089 and #6110 and follows up on the offline guard
introduced in #6184. Unlike the cached-repository case discussed there, this
issue only concerns inputs that already resolve to an explicit local file or
directory.

Expected behavior:

• An explicit local LoRA file should work in offline mode.
• An explicit local directory should support discovery of its LoRA weight file.
• A remote repository ID without weight_name should continue to raise in
  offline mode.
• model_info() should not be called in offline mode.

I am willing to submit a focused PR that moves the local file and directory
checks before the offline-mode guard and adds regression tests using temporary
local files without downloading any models.

Would this distinction and proposed scope be acceptable for a PR?

Reproduction

HF_HUB_OFFLINE=1

Run this in a fresh Python process:

import os

os.environ["HF_HUB_OFFLINE"] = "1"

import tempfile
from pathlib import Path

import torch
from safetensors.torch import save_file

from diffusers.loaders import StableDiffusionLoraLoaderMixin


with tempfile.TemporaryDirectory() as tmpdir:
    weight_path = Path(tmpdir) / "adapter.safetensors"
    save_file(
        {"unet.test.lora_A.weight": torch.ones(1)},
        str(weight_path),
    )

    StableDiffusionLoraLoaderMixin.lora_state_dict(tmpdir)

local_files_only=True

The same failure occurs with HF_HUB_OFFLINE disabled and only
local_files_only=True. Run this separately in a fresh Python process:

import os

os.environ.pop("HF_HUB_OFFLINE", None)

import tempfile
from pathlib import Path

import torch
from safetensors.torch import save_file

from diffusers.loaders import StableDiffusionLoraLoaderMixin


with tempfile.TemporaryDirectory() as tmpdir:
    weight_path = Path(tmpdir) / "adapter.safetensors"
    save_file(
        {"unet.test.lora_A.weight": torch.ones(1)},
        str(weight_path),
    )

    StableDiffusionLoraLoaderMixin.lora_state_dict(
        tmpdir,
        local_files_only=True,
    )

Both examples use a small temporary safetensors file and do not download any
models.

Logs

Traceback (most recent call last):
  File "<stdin>", line 21, in <module>
  File ".../huggingface_hub/utils/_validators.py", line 114, in _inner_fn
    return fn(*args, **kwargs)
  File ".../diffusers/loaders/lora_pipeline.py", line 318, in lora_state_dict
    state_dict, metadata = _fetch_state_dict(
  File ".../diffusers/loaders/lora_base.py", line 224, in _fetch_state_dict
    weight_name = _best_guess_weight_name(
  File ".../diffusers/loaders/lora_base.py", line 281, in _best_guess_weight_name
    raise ValueError("When using the offline mode, you must specify a `weight_name`.")
ValueError: When using the offline mode, you must specify a `weight_name`.

System Info

• 🤗 Diffusers version: 0.40.0.dev0
• Reproduced on main commit: 1aadc65
• Platform: Linux-5.4.210-39.1.pagevecsize-x86_64-with-glibc2.39
• Running on Google Colab?: No
• Python version: 3.12.3
• PyTorch version (GPU?): 2.6.0+cpu (False)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.36.2
• Transformers version: 4.57.6
• Accelerate version: 1.14.0
• PEFT version: 0.17.1
• Safetensors version: 0.8.0
• xFormers version: not installed
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

Who can help?

@sayakpaul