TypeError in rope validation: set -= list when config loaded from JSON

[Fr0do]

Bug description

_check_received_keys in modeling_rope_utils.py (line 919) performs received_keys -= ignore_keys where received_keys is a set but ignore_keys can be a list after JSON deserialization, causing:

TypeError: unsupported operand type(s) for -=: 'set' and 'list'

Root cause

Model configs like Qwen3.5 define ignore_keys_at_rope_validation = {"mrope_section", "mrope_interleaved"} as a Python set. However, JSON has no set type — when the config is serialized to JSON and loaded back (e.g. via huggingface_hub dataclass validation), sets become lists. The _check_received_keys method doesn't account for this.

Reproduction

# transformers==5.4.0, huggingface_hub>=1.7
from transformers import AutoConfig
# Works fine (loads from Python class):
config = AutoConfig.from_pretrained("Qwen/Qwen3.5-35B-A3B")

# Fails when loaded via huggingface_hub strict dataclass validation
# (e.g. through vLLM which triggers this path):
# StrictDataclassClassValidationError: Class validation error for validator 'validate_rope':
#     TypeError: unsupported operand type(s) for -=: 'set' and 'list'

Triggered by vLLM 0.18.0 serving Qwen/Qwen3.5-35B-A3B.

Affected models

Any model defining ignore_keys_at_rope_validation: Qwen3.5, ErnieVLMoe, GLM4V, Ministral3, etc.

Fix

# src/transformers/modeling_rope_utils.py, line 919
- received_keys -= ignore_keys
+ received_keys -= set(ignore_keys)

set() is a no-op when input is already a set, and correctly handles the list case.

Environment

• transformers 5.4.0
• huggingface_hub 1.7.2 / 1.8.0
• vLLM 0.18.0
• Python 3.13