transformers

[Qwen3Next] preserve linear-attn-mask optimization under torch.compile/export

#46148

Open