[CPU] Add mrope kernel for Qwen3-vl (#12531)

Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

sgl-kernel/csrc/cpu/torch_extension_cpu.cpp

@@ -306,6 +306,17 @@ std::tuple<at::Tensor, at::Tensor> rotary_embedding_cpu(
     at::Tensor& cos_sin_cache,
     bool is_neox);
 
+// mrope
+std::tuple<at::Tensor, at::Tensor> multimodal_rotary_embedding_cpu(
+    at::Tensor& positions,
+    at::Tensor& query,
+    at::Tensor& key,
+    int64_t head_size,
+    at::Tensor& cos_sin_cache,
+    const std::optional<std::vector<int64_t>>& mrope_section,
+    bool mrope_interleaved,
+    bool is_neox);
+
 // CPU and memory binding
 std::string init_cpu_threads_env(const std::string& cpu_ids);
 
@@ -518,6 +529,11 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
       "rotary_embedding_cpu(Tensor positions, Tensor query, Tensor key, int head_size, Tensor cos_sin_cache, "
       "bool is_neox) -> (Tensor, Tensor)");
   m.impl("rotary_embedding_cpu", torch::kCPU, &rotary_embedding_cpu);
+  // multimodal rope
+  m.def(
+      "multimodal_rotary_embedding_cpu(Tensor positions, Tensor query, Tensor key, int head_size, Tensor "
+      "cos_sin_cache, int[]? mrope_section, bool mrope_interleaved, bool is_neox) -> (Tensor, Tensor)");
+  m.impl("multimodal_rotary_embedding_cpu", torch::kCPU, &multimodal_rotary_embedding_cpu);
 
   // CPU and memory binding
   m.def("init_cpu_threads_env(str cpu_ids) -> str");