[CPU] Support chunk_gated_delta_rule kernel for Qwen3-Next (#12441)

sgl-kernel/csrc/cpu/torch_extension_cpu.cpp

@@ -100,6 +100,20 @@ void extend_attention_cpu(
     double sm_scale,
     double logit_cap);
 
+// linear attention
+std::tuple<at::Tensor, at::Tensor> chunk_gated_delta_rule_cpu(
+    const at::Tensor& query,
+    const at::Tensor& key,
+    const at::Tensor& value,
+    const at::Tensor& g,
+    const at::Tensor& beta,
+    const at::Tensor& initial_state,
+    bool output_final_state,
+    const at::Tensor& cu_seqlens,
+    bool head_first,
+    bool use_qk_l2norm_in_kernel,
+    double eps = 1e-5);
+
 // weight prepack
 at::Tensor convert_weight_packed(at::Tensor& weight);
 
@@ -287,6 +301,13 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
       "extend_start_loc, int max_len_extend, float sm_scale, float logit_cap) -> ()");
   m.impl("extend_attention_cpu", torch::kCPU, &extend_attention_cpu);
 
+  // linear attn
+  m.def(
+      "chunk_gated_delta_rule_cpu(Tensor query, Tensor key, Tensor value, Tensor g, Tensor beta, "
+      "Tensor initial_state, bool output_final_state, Tensor cu_seqlens, bool head_first, "
+      "bool use_qk_l2norm_in_kernel, float eps=1e-5) -> (Tensor, Tensor)");
+  m.impl("chunk_gated_delta_rule_cpu", torch::kCPU, &chunk_gated_delta_rule_cpu);
+
   // weight prepack
   m.def("convert_weight_packed(Tensor weight) -> Tensor");
   m.impl("convert_weight_packed", torch::kCPU, &convert_weight_packed);