Fix transformers 5.2.0 transpose conversion issue

[lvliang-intel]

Description

Root cause
In transformers 5.2.0, the new Transpose conversion class calls get_parameter() on weight tensors during quantized MoE checkpoint loading. auto_round's QuantLinear stores weight_packed and weight_scale as buffers (via register_buffer()), not nn.Parameter, so get_parameter() raises an AttributeError, which transformers then raises as a RuntimeError.

Fix
A _patch_transpose_for_buffers() monkey-patch wraps Transpose.convert() to check if the target tensor is a buffer first — if so, it returns the tensor as-is instead of attempting a transpose. This patch is applied in monkey_patch_transformers() for transformers ≥ 5.2.0.

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Code refactoring
• Other (please specify):

Related Issues

Fixes or relates to #

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.

[Copilot]

Pull request overview

Adds a compatibility monkey-patch to auto_round to prevent failures when loading quantized MoE checkpoints with transformers==5.2.0+, where a new transpose conversion path incorrectly assumes weights are nn.Parameter rather than buffers.

Changes:

• Introduce _patch_transpose_for_buffers() to wrap transformers.core_model_loading.Transpose.convert() and bypass transpose logic for buffer-backed tensors.
• Apply the patch from monkey_patch_transformers() for transformers >= 5.2.0.