[CPU] Sparse weights decompression: fixed property name and primitive caching

src/bindings/python/tests/test_runtime/test_properties.py

@@ -195,7 +195,7 @@ def test_properties_ro(ov_property_ro, expected_value):
         ),
         (
             properties.intel_cpu.sparse_weights_decompression_rate,
-            "SPARSE_WEIGHTS_DECOMPRESSION_RATE",
+            "CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE",
             (
                 (0.1, np.float32(0.1)),
                 (2.0, 2.0),

src/inference/include/openvino/runtime/intel_cpu/properties.hpp

@@ -47,7 +47,7 @@ namespace intel_cpu {
  */
 static constexpr Property<bool> denormals_optimization{"CPU_DENORMALS_OPTIMIZATION"};
 
-static constexpr Property<float> sparse_weights_decompression_rate{"SPARSE_WEIGHTS_DECOMPRESSION_RATE"};
+static constexpr Property<float> sparse_weights_decompression_rate{"CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE"};
 
 }  // namespace intel_cpu
 }  // namespace ov

src/plugins/intel_cpu/src/nodes/fullyconnected.cpp

@@ -591,7 +591,7 @@ void FullyConnected::createDescriptorInternal(const dnnl::memory::desc &inputDes
     dnnl::memory::desc wgh_candidate;
     if (useSparseWeights) {
         wgh_candidate = { DnnlExtensionUtils::convertToDnnlDims(getInputShapeAtPort(WEIGHTS_ID).getStaticDims()),
-                wdt, memory::desc::packed(nnzCount) };
+                wdt, memory::desc::packed() };
     } else {
         wgh_candidate = { DnnlExtensionUtils::convertToDnnlDims(getInputShapeAtPort(WEIGHTS_ID).getStaticDims()),
                                         wdt, dnnl::memory::format_tag::any };
@@ -930,10 +930,9 @@ bool FullyConnected::useSparseWeightsDecompression() {
             zerosCounts++;
         }
     }
-    nnzCount = elementsCount - zerosCounts;
 
     DEBUG_LOG(getName(), ", weightsData.size() = ", elementsCount, ", zerosCounts = ",
-        zerosCounts, ", nnzCount = ", nnzCount);
+        zerosCounts, ", nnzCount = ", elementsCount - zerosCounts);
 
     weiSparseRate = static_cast<float>(zerosCounts) / static_cast<float>(elementsCount);
 

src/plugins/intel_cpu/src/nodes/fullyconnected.h

@@ -42,7 +42,6 @@ class FullyConnected : public Node {
 
     void initSupportedPrimitiveDescriptors() override;
     void initOptimalPrimitiveDescriptor() override;
-    // void createPrimitive() override;
     std::shared_ptr<MemoryDesc> getSrcMemDesc(dnnl::primitive_desc_iterator &primitive_desc_it, size_t idx) override;
     std::shared_ptr<MemoryDesc> getDstMemDesc(dnnl::primitive_desc_iterator &primitive_desc_it, size_t idx) override;
 
@@ -112,7 +111,6 @@ class FullyConnected : public Node {
 
     // sparse weights
     bool useSparseWeights = false;
-    int nnzCount = -1;
     float minSparseRate = 1.f;
     float weiSparseRate = 0.f;
     bool useSparseWeightsDecompression();

src/plugins/intel_cpu/tests/functional/shared_tests_instances/behavior/ov_executable_network/get_metric.cpp

@@ -9,6 +9,7 @@
 #include "openvino/runtime/core.hpp"
 #include "openvino/runtime/compiled_model.hpp"
 #include "openvino/runtime/properties.hpp"
+#include "openvino/runtime/intel_cpu/properties.hpp"
 
 #include <gtest/gtest.h>
 
@@ -113,6 +114,13 @@ TEST_F(OVClassConfigTestCPU, smoke_CheckModelStreamsHasHigherPriorityThanThrough
     ASSERT_EQ(streams, value);
 }
 
+TEST_F(OVClassConfigTestCPU, smoke_CheckSparseWeigthsDecompressionRate) {
+    ov::Core core;
+
+    core.set_property(deviceName, ov::intel_cpu::sparse_weights_decompression_rate(0.8));
+    ASSERT_NO_THROW(ov::CompiledModel compiledModel = core.compile_model(model, deviceName));
+}
+
 const std::vector<ov::AnyMap> multiDevicePriorityConfigs = {
         {ov::device::priorities(CommonTestUtils::DEVICE_CPU)}};