Fix TypeError in save_pretrained error handling (fixes #38422)

[rahulrshetty45]

Summary

Fixes a TypeError in the save_pretrained error handling routine that occurs when shared_names contains shared tensors.

Problem

In src/transformers/modeling_utils.py line 3747, the code attempts to call set(shared_names) where shared_names is List[Set[str]]. This raises:

TypeError: unhashable type: 'set'

The bug occurs because you can't create a set from a list of sets (sets are unhashable).

Solution

Replace error_names.append(set(shared_names)) with error_names.extend(shared_names) to properly handle the List[Set[str]] structure.

This preserves the intended behavior where error_names is a list of sets, with each set representing a group of tensors that share memory/storage.

Testing

• Verified syntax and imports work correctly
• Tested the logic with sample data
• Confirmed backward compatibility
• Single line change with minimal impact

Related Issue

Fixes #38422

Changes

• src/transformers/modeling_utils.py: Replace append(set(shared_names)) with extend(shared_names)

[Rocketknight1]

Yes, this fix LGTM now! Thank you for the contribution and for iterating on the original

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.