Fix bug of referencing a temporary variable. (#14614)

test=develop

paddle/fluid/inference/api/analysis_predictor.cc

@@ -284,6 +284,7 @@ bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
         framework::GetFetchVariable(*scope, "fetch", idx);
     auto type = fetch.type();
     auto output = &(outputs->at(i));
+    output->name = fetchs_[idx]->Input("X")[0];
     if (type == typeid(float)) {
       GetFetchOne<float>(fetch, output);
       output->dtype = PaddleDType::FLOAT32;

paddle/fluid/inference/api/analysis_predictor.h

@@ -109,7 +109,7 @@ class AnalysisPredictor : public PaddlePredictor {
   std::map<std::string, size_t> feed_names_;
   std::vector<framework::OpDesc *> fetchs_;
   // Memory buffer for feed inputs. The temporary LoDTensor will cause serious
-  // concurrency problems, so cache them.
+  // concurrency problems, wrong results and memory leak, so cache them.
   std::vector<framework::LoDTensor> feed_tensors_;
   details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
 

paddle/fluid/inference/api/api_impl.cc

@@ -185,8 +185,12 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                << inputs.size();
     return false;
   }
+
+  // Cache the inputs memory for better concurrency performance.
+  feed_tensors_.resize(inputs.size());
+
   for (size_t i = 0; i < inputs.size(); ++i) {
-    framework::LoDTensor input;
+    auto &input = feed_tensors_[i];
     framework::DDim ddim = framework::make_ddim(inputs[i].shape);
     void *input_ptr;
     if (inputs[i].dtype == PaddleDType::INT64) {
@@ -261,6 +265,7 @@ bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
         framework::GetFetchVariable(*scope, "fetch", idx);
     auto type = fetch.type();
     auto output = &(outputs->at(i));
+    output->name = fetchs_[idx]->Input("X")[0];
     if (type == typeid(float)) {
       GetFetchOne<float>(fetch, output);
       output->dtype = PaddleDType::FLOAT32;

paddle/fluid/inference/api/api_impl.h

@@ -69,6 +69,9 @@ class NativePaddlePredictor : public PaddlePredictor {
   std::vector<framework::OpDesc *> feeds_;
   std::map<std::string, size_t> feed_names_;
   std::vector<framework::OpDesc *> fetchs_;
+  // Memory buffer for feed inputs. The temporary LoDTensor will cause serious
+  // concurrency problems, wrong results and memory leak, so cache them.
+  std::vector<framework::LoDTensor> feed_tensors_;
   // Do not use unique_ptr, use parent scope to delete
   framework::Scope *sub_scope_{nullptr};
   details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;