[Question]: models/llm/qwen3_5.h 必须编译 NPU 头文件，而不是 common 的

[cding-nv]

❓ Describe the question

models/llm/qwen3_5.h:27      #include "qwen3_next.h"                        ← 必须编译
                    :44      Qwen3_5ForCausalLMImpl : public Qwen3NextForCausalLMImpl

models/llm/qwen3_next.h:22   #include "core/layers/npu_torch/qwen3_next_decoder_layer_impl.h"   ← NPU-only 目录
                       :40   layer::Qwen3NextDecoderLayerImpl                                    ← NPU-only 类

models/llm/qwen3_next_hybrid_base.h:35  #include "core/layers/npu_torch/qwen3_next_hybrid_decoder_layer_base.h"  ← NPU-only 目录

[cding-nv]

$ grep -Rn USE_NPU xllm/models/dit/pipelines/

xllm/models/dit/pipelines/pipeline_flux_base.h:18:#if defined(USE_NPU)
xllm/models/dit/pipelines/pipeline_wan_i2v.h:30:#if defined(USE_NPU)
xllm/models/dit/pipelines/pipeline_wan_i2v.h:39:#if defined(USE_NPU)
xllm/models/dit/pipelines/pipeline_wan_i2v.h:158:#if defined(USE_NPU)
xllm/models/dit/pipelines/pipeline_wan_i2v.h:175:#if defined(USE_NPU)
xllm/models/dit/pipelines/pipeline_wan_i2v.h:547:#if defined(USE_NPU)
xllm/models/dit/pipelines/pipeline_wan_i2v.h:563:        LOG(FATAL) << "Rolling load requires USE_NPU";
xllm/models/dit/pipelines/pipeline_wan_i2v.h:714:#if defined(USE_NPU)

[XuZhang99]

Currently, you can implement a new model/layer file for dcu, even if it might be a bit redundant.

We are working on pytorch-based model/layer impl and will deprecate libtorch-based model/layer, which maybe will take some months.

You can find more detail from this pr #1891