What does this PR do?

Fixes an issue where ModelMixin.from_pretrained() could fail to load a fully cached sharded checkpoint when HF_HUB_OFFLINE=1.

When loading sharded checkpoints, _get_checkpoint_shard_files() could still call the Hugging Face Hub metadata API unless local_files_only=True was explicitly passed. In offline mode this caused an OfflineModeIsEnabled error even though all required checkpoint files were already available in the local cache.

This change makes HF_HUB_OFFLINE imply local_files_only for sharded checkpoint loading, preventing unnecessary Hub metadata requests in offline mode.

Which issue does this fix?

Fixes #14952

Motivation and context

Users running Diffusers in offline environments should be able to load cached sharded checkpoints without requiring network access.

The existing local_files_only behavior already supported this use case when explicitly enabled. This change makes the behavior consistent with the HF_HUB_OFFLINE environment setting.

What are the changes?

• Treat HF_HUB_OFFLINE as equivalent to local_files_only=True in _get_checkpoint_shard_files().
• Add a regression test verifying that a cached sharded checkpoint can be loaded successfully while offline.
• Verify that the loaded offline model has the same parameters as the initially cached model.

Tests

Focused regression test:

• pytest tests/models/test_modeling_common.py::TestModelUtils::test_hf_hub_offline_with_sharded_checkpoint
  • 1 passed

Existing related test:

• pytest tests/models/test_modeling_common.py::TestModelUtils::test_local_files_only_with_sharded_checkpoint
  • 1 passed

Also verified:

• git diff --check passes.

Final self-review

• The change is limited to the reported issue.
• The implementation uses the existing HF_HUB_OFFLINE behavior rather than introducing a new configuration mechanism.
• A regression test was added for the reported offline sharded-checkpoint scenario.
• The test verifies that model parameters remain identical when loading from the cache in offline mode.
• Existing related sharded-checkpoint behavior remains covered by its existing test.
• No unrelated files or formatting changes were introduced.
• The commit is signed off with DCO.