Revert "Fix TRT EP allocator memory leak (#16552)"

This reverts commit d8792f8.
microsoft · Aug 7, 2023 · dd6b443 · dd6b443
1 parent 6c604c1
commit dd6b443
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 17 deletions.
diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
@@ -14,7 +14,6 @@
 #include "core/providers/cuda/shared_inc/cuda_call.h"
 #include "core/providers/cuda/math/unary_elementwise_ops_impl.h"
 #include "core/providers/cuda/gpu_data_transfer.h"
-#include "core/session/allocator_adapters.h"
 #include "cuda_runtime_api.h"
 #include "core/common/gsl.h"
 #include <unordered_map>
@@ -992,12 +991,6 @@ TensorrtExecutionProvider::~TensorrtExecutionProvider() {
     ORT_IGNORE_RETURN_VALUE(CUDA_CALL(cudaStreamDestroy(stream_)));
   }
   ReleaseTensorRTCustomOpDomainList(info_.custom_op_domain_list);
-
-  if (alloc_ != nullptr) {
-    // This code is same as OrtApis::ReleaseAllocator defined in allocator_adapters.cc.
-    // We can't get api inside destructor so that's why we duplicate the code here.
-    delete static_cast<OrtAllocatorImpl*>(alloc_);
-  }
 }
 
 bool TensorrtExecutionProvider::IsGraphCaptureEnabled() const {
@@ -2266,18 +2259,15 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<FusedNodeAnd
       auto trt_context = trt_state->context->get();
       auto trt_profiles = trt_state->profiles;
       auto max_context_mem_size_ptr = trt_state->max_context_mem_size_ptr;
+      OrtMemoryInfo mem_info("", OrtAllocatorType::OrtDeviceAllocator, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, 0));
+      OrtAllocator* alloc;
+      Ort::ThrowOnError(api->KernelContext_GetAllocator(context, &mem_info, &alloc));
       int num_inputs = static_cast<int>(input_indexes.size());
       int num_outputs = static_cast<int>(output_indexes.size());
       bool engine_update = false;
       std::unordered_set<std::string> input_names;
       std::unordered_map<std::string, std::vector<int32_t>> tensor_shape_values;
 
-      OrtMemoryInfo mem_info("", OrtAllocatorType::OrtDeviceAllocator, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, 0));
-      if (alloc_ == nullptr) {
-        Ort::ThrowOnError(api->KernelContext_GetAllocator(context, &mem_info, &alloc_));
-      }
-      OrtAllocator* alloc = alloc_;
-
       void* cuda_stream;
       Ort::ThrowOnError(api->KernelContext_GetGPUComputeStream(context, &cuda_stream));
       cudaStream_t stream = static_cast<cudaStream_t>(cuda_stream);

diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.h b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.h
@@ -223,10 +223,6 @@ class TensorrtExecutionProvider : public IExecutionProvider {
   bool detailed_build_log_ = false;
   bool cuda_graph_enable_ = false;
 
-  // The OrtAllocator object will be get during ep compute time
-  // and should be kept for the lifetime of TRT EP object.
-  OrtAllocator* alloc_ = nullptr;
-
   std::unique_ptr<CUDAGraph> cuda_graph_;  // ORT TRT only supports CUDA graph when whole model is supported by TRT, so simply maintaining a CUDAGraph pointer is enough (no need to maintain one CUDAGraph pointer per TRT subgraph)
   bool is_graph_captured_ = false;
   int regular_run_count_before_graph_capture_ = 0;