[upstream] Expecting future huggingface/transformer incompatibility

[mratsim]

Hello team,

I was looking into adding my first MoECalibrationModule for a currently unsupported model (GLM4.5-Air, to prepare for upcoming GLM4.6-Air and also GLM4.6V) as I expect my previous quant quality was not up to par (though can't test the KL-divergence #2031).

However as luck would have it, https://github.com/huggingface/transformer seems to have merged a breaking 8.7K lines PR just yesterday: huggingface/transformers#41580

For example here is the new definition of Qwen3 MoE:

The gate has been removed from Qwen3MoeSparseBlock and moved to a Qwen3MoeTopKRouter

This departs from the current implementation in llmcompressor

Though I think we can just register both as calibration target and it should work ™️.

Note that the Hugging face PR is marked as tentative for a v5 of transformers so we might be OK with the old scheme for a time.

[kylesayrs]

@mratsim

Yes! Good catch. These changes have been in a pipeline for a while, and will dramatically improve LLM Compressor's ability to handle MoEs in a more standardized way. I agree, these changes will break current implementations.