Fix for blob being freed at the end of core.impot_model function

MirceaDan99 · Dec 10, 2024 · 6b5a462 · 6b5a462
1 parent e2c1b93
commit 6b5a462
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 3 deletions.
diff --git a/src/core/dev_api/openvino/runtime/shared_buffer.hpp b/src/core/dev_api/openvino/runtime/shared_buffer.hpp
@@ -12,6 +12,12 @@ namespace ov {
 template <typename T>
 class SharedBuffer : public ov::AlignedBuffer {
 public:
+
+    SharedBuffer(const SharedBuffer&) = delete;
+    SharedBuffer(SharedBuffer&&) = default;
+    SharedBuffer& operator=(const SharedBuffer&) = delete;
+    SharedBuffer& operator=(SharedBuffer&&) = default;
+
     SharedBuffer(char* data, size_t size, const T& shared_object) : _shared_object(shared_object) {
         m_allocated_buffer = data;
         m_aligned_buffer = data;

diff --git a/src/plugins/intel_npu/src/al/src/config/runtime.cpp b/src/plugins/intel_npu/src/al/src/config/runtime.cpp
@@ -7,7 +7,9 @@
 #include <sstream>
 
 #include "intel_npu/config/common.hpp"
+#include "openvino/util/mmap_object.hpp"
 #include "openvino/runtime/properties.hpp"
+#include "openvino/runtime/shared_buffer.hpp"
 
 using namespace intel_npu;
 using namespace ov::intel_npu;
@@ -158,16 +160,22 @@ std::string intel_npu::WORKLOAD_TYPE::toString(const ov::WorkloadType& val) {
 }
 
 //
-// WORKLOAD_TYPE
+// CACHED_MODEL_BUFFER
 //
 
 std::shared_ptr<ov::AlignedBuffer> intel_npu::CACHED_MODEL_BUFFER::parse(std::string_view val) {
     std::istringstream ss = std::istringstream(std::string(val));
     void* modelBufferPtr;
 
     ss >> modelBufferPtr;
-
-    return std::shared_ptr<ov::AlignedBuffer>(static_cast<ov::AlignedBuffer*>(modelBufferPtr));
+    // If we don't "steal" below resources from the casted ov::AlignedBuffer, parsed blob will be freed
+    // after core.import_model causing problems at inference.get_profiling_info()
+    auto* modelBufferSO = dynamic_cast<ov::SharedBuffer<std::shared_ptr<ov::MappedMemory>>*>(static_cast<ov::AlignedBuffer*>(modelBufferPtr));
+    std::shared_ptr<ov::MappedMemory> mappedMemorySOPtr;
+    auto modelBufferSOPtr = std::make_shared<ov::SharedBuffer<std::shared_ptr<ov::MappedMemory>>>(nullptr, 0, mappedMemorySOPtr);
+    *modelBufferSOPtr = std::move(*modelBufferSO);
+
+    return modelBufferSOPtr;
 }
 
 std::string intel_npu::CACHED_MODEL_BUFFER::toString(const std::shared_ptr<ov::AlignedBuffer>& val) {