transformers

Fix MXFP4 quantizer to support variable num_local_experts and hidden_size

#41795

Merged