bugfix: resolve compilation issues when building without NPU TORCH.

Author: yingxudeng

CMakeLists.txt

@@ -330,7 +330,7 @@ else()
 endif()
 
 if(USE_NPU)
-  add_definitions(-DUSE_NPU_TORCH)
+  # add_definitions(-DUSE_NPU_TORCH)
   add_definitions(-DUSE_NPU)
   add_definitions(-DBUILD_LIBTORCH)
   add_definitions(-DTORCH_SETCUSTOMHANDLER=ON)

xllm/models/llm/qwen3.h

@@ -45,7 +45,7 @@ class QWen3ModelImpl : public LlmModelImplBase<QWen3DecoderLayer> {
         xllm::layer::RmsNorm(
             model_args.hidden_size(), model_args.rms_norm_eps(), options));
 #else
-    norm_ = register_module("norm", layer::RmsNorm(context));
+    norm_ = register_module("norm", layer::NpuRmsNorm(context));
 #endif
     for (auto i = 0; i < FLAGS_micro_batch_num; i++) {
 #if defined(USE_NPU_TORCH)

xllm/models/llm/qwen3_moe.h

@@ -367,10 +367,13 @@ class Qwen3MoeModelImpl : public torch::nn::Module {
   torch::Dtype dtype_;
   layer::WordEmbedding embed_tokens_{nullptr};
   layer::AttentionMask attn_mask_;
-  layer::RmsNorm norm_{nullptr};
+
 #if defined(USE_NPU)
   torch::Tensor cos_sin_;
   layer::PosEmbedding atb_pos_emb_{nullptr};
+  layer::NpuRmsNorm norm_{nullptr};
+#else
+  layer::RmsNorm norm_{nullptr};
 #endif
   std::vector<int64_t> mrope_section_;
 };