!29283 [MSLITE] fix mem leak and diverse wide deep networks compatibility in tensorrt delegate

Merge pull request !29283 from Liu_Xuu/trt_0118_fullfusion

mindspore/lite/src/delegate/parameter_cache/load_host_cache_model.cc

@@ -29,6 +29,13 @@ constexpr size_t kGatherInputsSize = 3;
 }
 namespace mindspore {
 namespace cache {
+HostCacheModel::~HostCacheModel() {
+  if (cache_model_ != nullptr) {
+    delete cache_model_;
+    MS_LOG(ERROR) << "delete cache_model_";
+    cache_model_ = nullptr;
+  }
+}
 MSTensor *SchemaTensorToMSTensor(lite::SchemaTensorWrapper *schema_tensor_wrapper,
                                  mindspore::schema::Tensor *schema_tensor) {
   std::vector<int64_t> shape;

mindspore/lite/src/delegate/parameter_cache/load_host_cache_model.h

@@ -30,7 +30,8 @@ namespace mindspore {
 namespace cache {
 class HostCacheModel {
  public:
-  HostCacheModel() {}
+  HostCacheModel() = default;
+  ~HostCacheModel();
   Status LoadCache(const std::string &model_path);
   Status LoadCache(DelegateModel<schema::Primitive> *model);
   bool CheckIsCacheKernel(kernel::Kernel *kernel);

mindspore/lite/src/delegate/tensorrt/op/matmul_tensorrt.cc

@@ -49,6 +49,12 @@ int MatMulTensorRT::AddInnerOp(nvinfer1::INetworkDefinition *network) {
     transpose_b_ = primitive->transpose_b() ? nvinfer1::MatrixOperation::kTRANSPOSE : nvinfer1::MatrixOperation::kNONE;
     activation_ = primitive->activation_type();
   } else if (type_ == schema::PrimitiveType_FullConnection) {
+    auto primitive = this->GetPrimitive()->value_as_FullConnection();
+    if (primitive == nullptr) {
+      MS_LOG(ERROR) << "convert to primitive FullConnection failed for " << op_name_;
+      return RET_ERROR;
+    }
+    activation_ = primitive->activation_type();
     transpose_a_ = nvinfer1::MatrixOperation::kNONE;
     transpose_b_ = nvinfer1::MatrixOperation::kTRANSPOSE;
   }

mindspore/lite/src/delegate/tensorrt/tensorrt_serializer.cc

@@ -36,6 +36,7 @@ nvinfer1::ICudaEngine *TensorRTSerializer::GetSerializedEngine() {
     return nullptr;
   }
   nvinfer1::ICudaEngine *engine = runtime->deserializeCudaEngine(trt_model_stream, size, nullptr);
+  delete[] trt_model_stream;
   runtime->destroy();
   return engine;
 }

mindspore/lite/src/lite_session.cc

@@ -940,6 +940,8 @@ int LiteSession::DelegateInit() {
 int LiteSession::Init(InnerContext *context) {
   bool expected = false;
   if (!is_running_.compare_exchange_strong(expected, true)) {
+    delete context;
+    context = nullptr;
     MS_LOG(ERROR) << "Not support multi-threading";
     return RET_ERROR;
   }
@@ -1784,6 +1786,7 @@ int lite::LiteSession::LoadModelAndCompileByPath(const std::string &model_path,
   (reinterpret_cast<lite::LiteModel *>(model))->set_keep_model_buf(true);
   auto ret = CompileGraph(model);
   if (ret != lite::RET_OK) {
+    delete model;
     MS_LOG(ERROR) << "Compile model failed";
     return RET_ERROR;
   }
@@ -1808,6 +1811,7 @@ int lite::LiteSession::LoadModelAndCompileByPath(const std::string &model_path,
   (reinterpret_cast<lite::LiteModel *>(model))->set_keep_model_buf(true);
   auto ret = CompileGraph(model);
   if (ret != lite::RET_OK) {
+    delete model;
     MS_LOG(ERROR) << "Compile model failed";
     return RET_ERROR;
   }

mindspore/lite/src/scheduler.cc

@@ -515,7 +515,12 @@ int Scheduler::InitDelegateKernels(std::vector<kernel::LiteKernel *> *dst_kernel
     }
     auto ret = ReplaceDelegateKernels(&tmp_kernels);
     if (ret != RET_OK) {
-      MS_LOG(ERROR) << "NPU delegate replace delegate kernels failed.";
+      dst_kernels->insert(dst_kernels->end(), src_kernels.begin(), src_kernels.end());
+      dst_kernels->insert(dst_kernels->end(), tmp_kernels.begin(), tmp_kernels.end());
+      if (remain_kernel != nullptr) {
+        dst_kernels->push_back(remain_kernel);
+      }
+      MS_LOG(ERROR) << "Inner delegate replace delegate kernels failed.";
       return ret;
     }