[Paddle-TRT] tensorrt engine memory sharing default true (PaddlePaddl…

…e#58251)

* tensorrt engine memory sharing default true

* set to false when trt version less than 7.2

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc

@@ -586,6 +586,13 @@ std::string TensorRtSubgraphPass::CreateTensorRTOp(
   auto inspector_serialize = Get<bool>("inspector_serialize");
   auto disable_trt_plugin_fp16 = Get<bool>("disable_trt_plugin_fp16");
   auto context_memory_sharing = Get<bool>("context_memory_sharing");
+  if (context_memory_sharing && TRT_VERSION < 7200) {
+    // https://forums.developer.nvidia.com/t/nvinfer1-createexecutioncontextwithoutdevicememory-returns-nullptr/111878/2
+    // when trt version less than 7.2,
+    // createExecutionContextWithoutDeviceMemory() has bug.
+    // so, we cannot enable engine context memory sharing.
+    context_memory_sharing = false;
+  }
   auto enable_low_precision_io = Get<bool>("enable_low_precision_io");
   auto workspace_size = Get<int64_t>("workspace_size");
   auto gpu_device_id = Get<int>("gpu_device_id");

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -1291,7 +1291,7 @@ struct PD_INFER_DECL AnalysisConfig {
 
   // memory reuse related.
   bool enable_memory_optim_{false};
-  bool trt_engine_memory_sharing_{false};
+  bool trt_engine_memory_sharing_{true};
   int trt_engine_memory_sharing_identifier_{0};
 
   std::unordered_set<std::string> trt_ops_run_float_;