Revert changes in new import_model API, so NPU plugin will have `…

…unusedStream` when `model_buffer` is given
MirceaDan99 · Nov 27, 2024 · bc4ca08 · bc4ca08
1 parent a9f3271
commit bc4ca08
Show file tree

Hide file tree

Showing 9 changed files with 32 additions and 61 deletions.
diff --git a/src/inference/dev_api/openvino/runtime/iplugin.hpp b/src/inference/dev_api/openvino/runtime/iplugin.hpp
@@ -193,7 +193,8 @@ class OPENVINO_RUNTIME_API IPlugin : public std::enable_shared_from_this<IPlugin
      * @param properties A ov::AnyMap of properties
      * @return An Compiled model
      */
-    virtual std::shared_ptr<ov::ICompiledModel> import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+    virtual std::shared_ptr<ov::ICompiledModel> import_model(std::istream& model,
+                                                             std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                              const ov::AnyMap& properties) const;
 
     /**
@@ -206,7 +207,8 @@ class OPENVINO_RUNTIME_API IPlugin : public std::enable_shared_from_this<IPlugin
      * @param properties A ov::AnyMap of properties
      * @return An Compiled model
      */
-    virtual std::shared_ptr<ov::ICompiledModel> import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+    virtual std::shared_ptr<ov::ICompiledModel> import_model(std::istream& model,
+                                                             std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                              const ov::SoPtr<ov::IRemoteContext>& context,
                                                              const ov::AnyMap& properties) const;
 

diff --git a/src/inference/src/dev/core_impl.cpp b/src/inference/src/dev/core_impl.cpp
@@ -1460,8 +1460,8 @@ ov::SoPtr<ov::ICompiledModel> ov::CoreImpl::load_model_from_cache(
                     }
                 }
                 if (model_buffer) {
-                    compiled_model = context ? plugin.import_model(model_buffer, context, update_config)
-                                             : plugin.import_model(model_buffer, update_config);
+                    compiled_model = context ? plugin.import_model(networkStream, model_buffer, context, update_config)
+                                             : plugin.import_model(networkStream, model_buffer, update_config);
                 } else {
                     compiled_model = context ? plugin.import_model(networkStream, context, update_config)
                                              : plugin.import_model(networkStream, update_config);

diff --git a/src/inference/src/dev/iplugin.cpp b/src/inference/src/dev/iplugin.cpp
@@ -57,12 +57,14 @@ const std::string& ov::IPlugin::get_device_name() const {
     return m_plugin_name;
 }
 
-std::shared_ptr<ov::ICompiledModel> ov::IPlugin::import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+std::shared_ptr<ov::ICompiledModel> ov::IPlugin::import_model(std::istream& model,
+                                                              std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                               const ov::AnyMap& properties) const{
     OPENVINO_THROW_NOT_IMPLEMENTED("This method is not implemented");
 }
 
-std::shared_ptr<ov::ICompiledModel> ov::IPlugin::import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+std::shared_ptr<ov::ICompiledModel> ov::IPlugin::import_model(std::istream& model,
+                                                              std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                               const ov::SoPtr<ov::IRemoteContext>& context,
                                                               const ov::AnyMap& properties) const{
     OPENVINO_THROW_NOT_IMPLEMENTED("This method is not implemented");

diff --git a/src/inference/src/dev/plugin.cpp b/src/inference/src/dev/plugin.cpp
@@ -79,14 +79,15 @@ ov::SoPtr<ov::ICompiledModel> ov::Plugin::import_model(std::istream& model,
     OV_PLUGIN_CALL_STATEMENT(return {m_ptr->import_model(model, context, config), m_so});
 }
 
-ov::SoPtr<ov::ICompiledModel> ov::Plugin::import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer, const ov::AnyMap& properties) const {
-    OV_PLUGIN_CALL_STATEMENT(return {m_ptr->import_model(model_buffer, properties), m_so});
+ov::SoPtr<ov::ICompiledModel> ov::Plugin::import_model(std::istream& model, std::shared_ptr<ov::AlignedBuffer> model_buffer, const ov::AnyMap& properties) const {
+    OV_PLUGIN_CALL_STATEMENT(return {m_ptr->import_model(model, model_buffer, properties), m_so});
 }
 
-ov::SoPtr<ov::ICompiledModel> ov::Plugin::import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+ov::SoPtr<ov::ICompiledModel> ov::Plugin::import_model(std::istream& model,
+                                                       std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                        const ov::SoPtr<ov::IRemoteContext>& context,
                                                        const ov::AnyMap& config) const {
-    OV_PLUGIN_CALL_STATEMENT(return {m_ptr->import_model(model_buffer, context, config), m_so});
+    OV_PLUGIN_CALL_STATEMENT(return {m_ptr->import_model(model, model_buffer, context, config), m_so});
 }
 
 ov::SoPtr<ov::IRemoteContext> ov::Plugin::create_context(const AnyMap& params) const {

diff --git a/src/inference/src/dev/plugin.hpp b/src/inference/src/dev/plugin.hpp
@@ -59,9 +59,10 @@ class Plugin {
                                            const ov::SoPtr<ov::IRemoteContext>& context,
                                            const ov::AnyMap& config) const;
 
-    SoPtr<ov::ICompiledModel> import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer, const ov::AnyMap& properties) const;
+    SoPtr<ov::ICompiledModel> import_model(std::istream& model, std::shared_ptr<ov::AlignedBuffer> model_buffer, const ov::AnyMap& properties) const;
 
-    SoPtr<ov::ICompiledModel> import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+    SoPtr<ov::ICompiledModel> import_model(std::istream& model,
+                                           std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                            const ov::SoPtr<ov::IRemoteContext>& context,
                                            const ov::AnyMap& config) const;
 

diff --git a/src/plugins/intel_cpu/src/plugin.cpp b/src/plugins/intel_cpu/src/plugin.cpp
@@ -553,48 +553,12 @@ ov::SupportedOpsMap Plugin::query_model(const std::shared_ptr<const ov::Model>&
 
 std::shared_ptr<ov::ICompiledModel> Plugin::import_model(std::istream& model_stream,
                                                          const ov::AnyMap& config) const {
-    OV_ITT_SCOPE(FIRST_INFERENCE, itt::domains::intel_cpu_LT, "import_model");
-
-    CacheDecrypt decrypt{ codec_xor };
-    bool decript_from_string = false;
-    if (config.count(ov::cache_encryption_callbacks.name())) {
-        auto encryption_callbacks = config.at(ov::cache_encryption_callbacks.name()).as<EncryptionCallbacks>();
-        decrypt.m_decrypt_str = encryption_callbacks.decrypt;
-        decript_from_string = true;
-    }
-
-    ModelDeserializer deserializer(
-        model_stream,
-        nullptr,
-        [this](const std::shared_ptr<ov::AlignedBuffer>& model, const std::shared_ptr<ov::AlignedBuffer>& weights) {
-            return get_core()->read_model(model, weights);
-        },
-        decrypt, decript_from_string);
-
-    std::shared_ptr<ov::Model> model;
-    deserializer >> model;
-
-    Config conf = engConfig;
-    Config::ModelType modelType = getModelType(model);
-
-    // check ov::loaded_from_cache property and erase it to avoid exception in readProperties.
-    auto _config = config;
-    const auto& it = _config.find(ov::loaded_from_cache.name());
-    bool loaded_from_cache = false;
-    if (it != _config.end()) {
-        loaded_from_cache = it->second.as<bool>();
-        _config.erase(it);
-    }
-    conf.readProperties(_config, modelType);
-
-    // import config props from caching model
-    calculate_streams(conf, model, true);
-    auto compiled_model = std::make_shared<CompiledModel>(model, shared_from_this(), conf, loaded_from_cache);
-    return compiled_model;
+    return import_model(model_stream, nullptr, config);
 }
 
 
-std::shared_ptr<ov::ICompiledModel> Plugin::import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+std::shared_ptr<ov::ICompiledModel> Plugin::import_model(std::istream& model_stream,
+                                                         std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                          const ov::AnyMap& config) const {
     OV_ITT_SCOPE(FIRST_INFERENCE, itt::domains::intel_cpu_LT, "import_model");
 
@@ -606,10 +570,8 @@ std::shared_ptr<ov::ICompiledModel> Plugin::import_model(std::shared_ptr<ov::Ali
         decript_from_string = true;
     }
 
-    std::stringstream empty_model_stream("");
-
     ModelDeserializer deserializer(
-        empty_model_stream,
+        model_stream,
         model_buffer,
         [this](const std::shared_ptr<ov::AlignedBuffer>& model, const std::shared_ptr<ov::AlignedBuffer>& weights) {
             return get_core()->read_model(model, weights);

diff --git a/src/plugins/intel_cpu/src/plugin.h b/src/plugins/intel_cpu/src/plugin.h
@@ -34,9 +34,11 @@ class Plugin : public ov::IPlugin {
             "import_model with RemoteContext is not supported by CPU plugin!");
     };
 
-    std::shared_ptr<ov::ICompiledModel> import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+    std::shared_ptr<ov::ICompiledModel> import_model(std::istream& model,
+                                                     std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                      const ov::AnyMap& properties) const override;
-    std::shared_ptr<ov::ICompiledModel> import_model(std::shared_ptr<ov::AlignedBuffer> model_buffer,
+    std::shared_ptr<ov::ICompiledModel> import_model(std::istream& model,
+                                                     std::shared_ptr<ov::AlignedBuffer> model_buffer,
                                                      const ov::SoPtr<ov::IRemoteContext>& context,
                                                      const ov::AnyMap& properties) const override {
         OPENVINO_THROW_NOT_IMPLEMENTED(

diff --git a/src/plugins/intel_cpu/src/utils/serialize.cpp b/src/plugins/intel_cpu/src/utils/serialize.cpp
@@ -47,13 +47,14 @@ ModelDeserializer::ModelDeserializer(std::istream& model_stream,
 
     void ModelDeserializer::operator>>(std::shared_ptr<ov::Model>& model) {
         if (m_model_buffer) {
-            process_mmap(m_model_buffer);
+            process_mmap(model, m_model_buffer);
         } else {
             process_stream(model);
         }
 }
 
-void ModelDeserializer::process_mmap(const std::shared_ptr<ov::AlignedBuffer>& mmemory) {
+void ModelDeserializer::process_mmap(std::shared_ptr<ov::Model>& model,
+                                     const std::shared_ptr<ov::AlignedBuffer>& mmemory) {
     // Note: Don't use seekg with mmaped stream. This may affect the performance of some models.
     // Get file size before seek content.
     // Blob from cache may have other header, so need to skip this.
@@ -106,7 +107,7 @@ void ModelDeserializer::process_mmap(const std::shared_ptr<ov::AlignedBuffer>& m
                                                                              hdr.model_size,
                                                                              xml_buff);
 
-    auto model = m_model_builder(model_buf, weights_buf);
+    model = m_model_builder(model_buf, weights_buf);
 
     // Set Info
     pugi::xml_node root = xml_in_out_doc.child("cnndata");

diff --git a/src/plugins/intel_cpu/src/utils/serialize.hpp b/src/plugins/intel_cpu/src/utils/serialize.hpp
@@ -31,7 +31,7 @@ class ModelDeserializer {
 public:
     typedef std::function<std::shared_ptr<ov::Model>(const std::shared_ptr<ov::AlignedBuffer>&, const std::shared_ptr<ov::AlignedBuffer>&)> ModelBuilder;
 
-    ModelDeserializer(std::istream& model_stream,
+    ModelDeserializer(std::istream& model,
                       std::shared_ptr<ov::AlignedBuffer> model_buffer,
                       ModelBuilder fn,
                       const CacheDecrypt& encrypt_fn,
@@ -44,7 +44,7 @@ class ModelDeserializer {
 protected:
     static void set_info(pugi::xml_node& root, std::shared_ptr<ov::Model>& model);
 
-    void process_mmap(const std::shared_ptr<ov::AlignedBuffer>& memory);
+    void process_mmap(std::shared_ptr<ov::Model>& model, const std::shared_ptr<ov::AlignedBuffer>& memory);
 
     void process_stream(std::shared_ptr<ov::Model>& model);