Cherry-pick pull-request for patch release 0.8.3 (#1592)

Author: baijumeswani

src/dml/interface.cpp

@@ -218,6 +218,17 @@ void InitDmlInterface(LUID* p_device_luid) {
     g_dml_device = std::make_unique<Dml::InterfaceImpl>(p_device_luid);
 }
 
+void CloseDmlInterface() {
+  Dml::dml_device_.Reset();
+  g_dml_device.reset();
+  Dml::dml_pooled_upload_heap_.reset();
+  Dml::dml_execution_context_.reset();
+  Dml::dml_readback_heap_.reset();
+  Dml::smart_directml_dll_.reset();
+  Dml::dml_objects_ = {};
+  Dml::ort_allocator_ = nullptr;
+}
+
 void SetDmlProvider(OrtSessionOptions& session_options) {
   Ort::ThrowOnError(Dml::dml_api_->SessionOptionsAppendExecutionProvider_DML1(&session_options, Dml::dml_device_.Get(), Dml::dml_objects_.command_queue.Get()));
 }

src/dml/interface.h

@@ -13,6 +13,8 @@ namespace Generators {
 void InitDmlInterface(LUID* p_device_luid);
 void SetDmlProvider(OrtSessionOptions& options);
 
+void CloseDmlInterface();
+
 DeviceInterface* GetDmlInterface();
 
 }  // namespace Generators

src/models/model.cpp

@@ -453,7 +453,6 @@ DeviceInterface* SetProviderSessionOptions(OrtSessionOptions& session_options,
 
       if (!disable_graph_capture) {
         session_options.AddConfigEntry("ep.dml.enable_graph_capture", "1");
-        session_options.AddConfigEntry("ep.dml.disable_memory_arena", "1");
       }
 
       SetDmlProvider(session_options);
@@ -643,7 +642,21 @@ Model::Model(std::unique_ptr<Config> config) : config_{std::move(config)} {
   p_device_kvcache_ = p_device_;
 }
 
-Model::~Model() = default;
+Model::~Model() {
+#if USE_DML
+  if (p_device_->GetType() == DeviceType::DML) {
+    auto& allocator = GetOrtGlobals()->device_allocators_[static_cast<int>(DeviceType::DML)];
+    allocator.session_.reset();
+    allocator.allocator_.reset();
+    session_options_.reset();
+    // DML objects are globally scoped and launch background threads that retain hardware resources.
+    // These threads persist beyond the lifetime of a Model, preventing proper cleanup and potentially causing deadlocks.
+    // To avoid blocking driver threads, we explicitly destroy DML objects when the Model is destroyed.
+    // They will be recreated as needed when a new Model is initialized.
+    CloseDmlInterface();
+  }
+#endif
+}
 
 void Model::CreateSessionOptionsFromConfig(const Config::SessionOptions& config_session_options,
                                            OrtSessionOptions& session_options,

src/tensor.cpp

@@ -16,6 +16,7 @@ Tensor::Tensor(std::unique_ptr<OrtValue> ort_tensor) : ort_tensor_{std::move(ort
 Tensor::~Tensor() {
   if (buffer_ != nullptr) {
     p_device_->GetAllocator().Free(buffer_);
+    buffer_ = nullptr;
   }
 }
 

tools/python/util/android.py

@@ -126,7 +126,7 @@ def start_emulator(
             "-no-audio",
             "-no-boot-anim",
             "-gpu",
-            "guest",
+            "swiftshader_indirect",
             "-delay-adb",
         ]