fix: handle removal of distributed_operation in Transformers TP rework

[peft-jambot]

Problem

Transformers PR #47579 reworked tensor parallelism (TP) to use a DTensor-based API, replacing the legacy distributed_operation mechanism. This PR removed the distributed_operation attribute from WeightTransform.__slots__ and __init__ in core_model_loading.py.

PEFT's build_peft_weight_mapping function in src/peft/utils/transformers_weight_conversion.py directly accessed orig_conversion.distributed_operation when building new WeightConverter objects for MoE LoRA adapter conversion (lines 279 and 321). With the new Transformers code, this attribute no longer exists, causing:

AttributeError: 'WeightConverter' object has no attribute 'distributed_operation'

This breaks loading LoRA adapters on MoE models (e.g. Mixtral) when using the new Transformers TP code, as exercised by the Transformers test test_mixtral_lora_conversion.

Solution

Introduced a _copy_runtime_attributes helper that copies optional runtime attributes (distributed_operation and quantization_operation) from the original converter to the new PEFT-specific converter using hasattr/setattr. Attributes that don't exist on the source (like distributed_operation on new Transformers) are silently skipped.

This is backward compatible:

• Older Transformers (with distributed_operation): The attribute is still copied as before.
• Newer Transformers (PR #47579, without distributed_operation): The attribute is skipped, and only quantization_operation is copied.

Testing

Ran the following tests with both Transformers main (v5.16.0.dev0, still has distributed_operation) and Transformers PR #47579 branch (v5.15.0.dev0, distributed_operation removed):

pytest tests/test_integrations.py -k "mixtral" -v

All 5 mixtral tests pass on both versions. Also ran the new regression test and the Transformers-side test:

pytest tests/test_integrations.py::TestTransformersV5::test_build_peft_weight_mapping_without_distributed_operation
pytest tests/peft_integration/test_peft_integration.py::PeftIntegrationTester::test_mixtral_lora_conversion

Both pass on both Transformers versions.

Regression test

Added test_build_peft_weight_mapping_without_distributed_operation to tests/test_integrations.py which constructs WeightConverter objects (simulating what get_model_conversion_mapping returns) and verifies build_peft_weight_mapping works regardless of whether distributed_operation exists.

Notes

• AI assistance was used for this PR.
• The fix is minimal and does not change any behavior for existing Transformers versions.