serialize kv-cache precision information for ScaledDotProductAttention

Author: liubo-intel

src/plugins/intel_cpu/src/graph_dumper.cpp

@@ -22,6 +22,7 @@
 
 #include "cpu_types.h"
 #include "node.h"
+#include "nodes/scaled_attn.h"
 #include "onednn/dnnl.h"
 #include "openvino/core/except.hpp"
 #include "openvino/core/model.hpp"
@@ -122,6 +123,13 @@ std::map<std::string, std::string> extract_node_metadata(const NodePtr& node) {
     serialization_info[ov::exec_model_info::EXECUTION_ORDER] = std::to_string(node->getExecIndex());
 
     serialization_info[ov::exec_model_info::RUNTIME_PRECISION] = node->getRuntimePrecision().get_type_name();
+    // record kv cache precision for ScaledDotProductAttention node
+    if (node->getType() == Type::ScaledDotProductAttention) {
+        auto* sdpa_node = dynamic_cast<ov::intel_cpu::node::ScaledDotProductAttention*>(node.get());
+        if (sdpa_node) {
+            serialization_info["kv_cache_precision"] = sdpa_node->getKVCachePrecision().get_type_name();
+        }
+    }
 
     return serialization_info;
 }