From 117b2e7743aea93fae5ac503195b230ac60d2b45 Mon Sep 17 00:00:00 2001
From: Matthew Hill <matthill@gmail.com>
Date: Wed, 10 Jun 2020 19:56:51 -0400
Subject: [PATCH] Fix GPU memory leak on TensorRT (#4172)

---
 .../core/providers/tensorrt/tensorrt_execution_provider.cc      | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
index e70105afa4..e721683ba9 100644
--- a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
+++ b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
@@ -1018,6 +1018,8 @@ common::Status TensorrtExecutionProvider::Compile(const std::vector<onnxruntime:
       for (int i = 0, end = num_binding_outputs; i < end; ++i) {
         if (output_types[i] == ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64) {
           cuda::Impl_Cast<int32_t, int64_t>(reinterpret_cast<int32_t*>(buffers[i + num_binding_inputs]), ort.GetTensorMutableData<int64_t>(output_tensor[i]), output_dim_sizes[i]);
+          cudaDeviceSynchronize();
+          cudaFree(buffers[i + num_binding_inputs]);
         }
       }