fix clang & apply review comments

Author: zhangYiIntel

src/common/transformations/include/transformations/common_optimizations/fuse_gated_delta_net.hpp

@@ -22,6 +22,14 @@ class TRANSFORMATIONS_API RemoveConcatSliceAfterLoop : public ov::pass::MatcherP
 /**
  * @ingroup ov_transformation_common_api
  * @brief Fuses a loop-based gated delta net sub-graph into an internal GatedDeltaNet operation.
+ *
+ * Expected Loop body semantics per step:
+ * 1) Decay recurrent state: `h_decay = h * exp(g)`
+ * 2) Compute projection and delta: `delta = v - reduce_sum(h_decay * k, axis=-2)`
+ * 3) Apply beta and update state: `h_new = h_decay + k * unsqueeze(delta * beta, axis=-2)`
+ * 4) Compute per-step output: `o = reduce_sum(h_new * q, axis=-2, keep_dims=true)` and scatter to time index
+ *
+ * The matcher validates this body shape/operation pattern before replacing the Loop with `GatedDeltaNet`.
  */
 
 class TRANSFORMATIONS_API FuseGDNLoop : public ov::pass::MatcherPass {

src/common/transformations/src/transformations/common_optimizations/fuse_gated_delta_net.cpp

@@ -71,7 +71,6 @@ bool matches_linear_attention_loop(const std::shared_ptr<ov::Node>& node) {
     auto step_index = any_input();
 
     auto step_index_unsqueeze = wrap_type<v0::Unsqueeze>({step_index, 0});
-    // auto recurrent_state_f32 = pattern::optional<v0::Convert>({recurrent_state});
     auto gate_f32 = pattern::optional<v0::Convert>({gate});
 
     auto exp_gate = wrap_type<v0::Exp>({gate_f32});
@@ -169,8 +168,11 @@ ov::pass::RemoveConcatSliceAfterLoop::RemoveConcatSliceAfterLoop() {
     auto slice_attn = pattern::wrap_type<ov::op::v8::Slice>({concat_loop, {0}, any_input(), {1}, {0}});
     auto reshape_attn = pattern::wrap_type<v1::Reshape>({slice_attn, any_input()},
                                                         pattern::shape_matches("[?, head_num, ?, v_head_size]"));
-    auto slice_state = pattern::wrap_type<ov::op::v8::Slice>({concat_loop, any_input(), {LLONG_MAX}, {1}, {0}});
-    auto reshape_state = pattern::wrap_type<v1::Reshape>({slice_state, any_input()});
+    auto state_end = pattern::wrap_type<v0::Constant>(value_matches("LLONG_MAX") || value_matches("-1"));
+    auto slice_state = pattern::wrap_type<ov::op::v8::Slice>({concat_loop, any_input(), state_end, {1}, {0}});
+    auto reshape_state =
+        pattern::wrap_type<v1::Reshape>({slice_state, any_input()},
+                                        pattern::shape_matches("[?, head_num, k_head_size, v_head_size]"));
     matcher_pass_callback callback = [OV_CAPTURE_CPY_AND_THIS](ov::pass::pattern::Matcher& m) {
         const auto& pattern_map = m.get_pattern_value_map();
         bool changed = false;

src/core/src/op/gated_delta_net.cpp

@@ -104,7 +104,7 @@ void GatedDeltaNet::validate_and_infer_types() {
                           "The head size in key and query should be the same, but got ",
                           k_head_size,
                           " and ",
-                          v_head_size,
+                          q_head_size,
                           ".");
 
     const auto gate_head_num = gate_ps[2];

src/plugins/intel_cpu/src/cpu_parallel.hpp

@@ -38,7 +38,7 @@ class CpuParallel {
                                                                          : parallel_get_max_threads() * m_multiplier;
         return num;
     }
-    [[nodiscard]] int get_num_worker_threads() const {
+    [[nodiscard]] static int get_num_worker_threads() {
         return parallel_get_max_threads();
     }
     void activate() const {

src/plugins/intel_cpu/src/nodes/gated_delta_net.cpp

@@ -14,6 +14,7 @@
 #include "graph_context.h"
 #include "kernels/linear_attn/recurrent_linear_attn.hpp"
 #include "memory_desc/cpu_blocked_memory_desc.h"
+#include "memory_desc/cpu_memory_desc.h"
 #include "node.h"
 #include "onednn/iml_type_mapper.h"
 #include "openvino/core/except.hpp"